def accelerate_filter(bag, p, state_vars, binders, args, cache, size):
parts = list(break_conj(p.body))
guards = []
map_conds = []
in_conds = []
others = []
for part in parts:
if p.arg not in free_vars(part):
others.append(part)
elif all((v == p.arg or v in state_vars) for v in free_vars(part)):
guards.append(part)
elif infer_map_lookup(part, p.arg, state_vars):
map_conds.append(part)
elif isinstance(part, EBinOp) and part.op == BOp.In and all(
v in state_vars for v in free_vars(part.e2)):
in_conds.append(part)
else:
others.append(part)
if in_conds:
for i in range(len(in_conds)):
rest = others + in_conds[:i] + in_conds[i + 1:] + map_conds
for tup in map_accelerate(
EFilter(
EFilter(bag,
ELambda(p.arg,
EAll(guards))).with_type(bag.type),
ELambda(p.arg, in_conds[i])).with_type(bag.type),
state_vars, binders, args, cache, size):
(e, pool) = tup
yield tup
if e.type == bag.type and pool == RUNTIME_POOL:
yield (EFilter(e, ELambda(p.arg,
EAll(rest))).with_type(bag.type),
RUNTIME_POOL)
def optimize_filter_as_if_distinct(xs, p, args, dnf=True):
if isinstance(xs, EBinOp) and xs.op == "-":
return EBinOp(optimize_filter_as_if_distinct(xs.e1, p, args), "-",
optimize_filter_as_if_distinct(xs.e2, p,
args)).with_type(xs.type)
# return optimize_filter_as_if_distinct(xs.e1, ELambda(p.arg, EAll([ENot(optimized_in(p.arg, xs.e2)), p.body])), args)
if dnf:
from cozy.syntax_tools import dnf, nnf
cases = dnf(nnf(p.body))
cases = [list(unique(c)) for c in cases]
# for c in cases:
# print("; ".join(pprint(x) for x in c))
assert cases
else:
cases = [list(break_conj(p.body))]
res = xs
for c in sorted(unique(cases[0]),
key=lambda c: 1
if any(v in args for v in free_vars(c)) else 0):
res = _simple_filter(res, ELambda(p.arg, c), args)
if cases[1:]:
cond = EAll([ENot(EAll(cases[0])), EAny(EAll(c) for c in cases[1:])])
res = EBinOp(
res, "+",
optimize_filter_as_if_distinct(xs,
ELambda(p.arg, cond),
args=args,
dnf=False)).with_type(res.type)
return res
def infer_map_lookup(filter, binder, state: {EVar}):
map_conds = []
other_conds = []
for c in break_conj(filter):
if list(infer_key_and_value(c, (binder, ), state)):
map_conds.append(c)
else:
other_conds.append(c)
if map_conds:
for (_, key_proj,
key_lookup) in infer_key_and_value(EAll(map_conds), (binder, ),
state):
return (key_proj, key_lookup, EAll(other_conds))
else:
return None
assert False
def improve(target: Exp,
context: Context,
assumptions: Exp = T,
stop_callback=never_stop,
hints: [Exp] = (),
examples: [{
str: object
}] = (),
cost_model: CostModel = 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`.
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})""".format(target=target,
context=context,
assumptions=assumptions,
stop_callback=stop_callback,
hints=hints,
examples=examples,
cost_model=cost_model))
target = freshen_binders(target, context)
assumptions = freshen_binders(assumptions, context)
print()
print("improving: {}".format(pprint(target)))
print("subject to: {}".format(pprint(assumptions)))
print()
try:
assert exp_wf(target, context=context, assumptions=assumptions)
except ExpIsNotWf as ex:
print(
"WARNING: initial target is not well-formed [{}]; this might go poorly..."
.format(str(ex)))
print(pprint(ex.offending_subexpression))
print(pprint(ex.offending_subexpression.type))
# raise
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 = None
if incremental.value:
solver = IncrementalSolver(vars=vars, funcs=funcs)
solver.add_assumption(assumptions)
_sat = solver.satisfy
else:
_sat = lambda e: satisfy(e, vars=vars, funcs=funcs)
if _sat(assumptions) is None:
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, assumptions, context, examples,
cost_model, stop_callback, hints)
try:
while True:
# 1. find any potential improvement to any sub-exp of target
for new_target in learner.next():
print("Found candidate improvement: {}".format(
pprint(new_target)))
# 2. check
with task("verifying candidate"):
if incremental.value:
solver.push()
solver.add_assumption(
ENot(
EBinOp(target, "==",
new_target).with_type(BOOL)))
counterexample = _sat(T)
else:
formula = EAll([
assumptions,
ENot(
EBinOp(target, "==",
new_target).with_type(BOOL))
])
counterexample = _sat(formula)
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")
raise NoMoreImprovements()
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
if incremental.value:
solver.pop()
except NoMoreImprovements:
return
except KeyboardInterrupt:
raise
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