lean2/src/kernel/converter.cpp
Leonardo de Moura e206fcc1af perf(kernel/type_checker): reduce the overhead of creating delayed_justification objects, a huge number of them is created when type checking applications
We reduce the cost by avoiding the allocation of std::functional objects, and the unnecessary increment/decrement of reference counters.

Signed-off-by: Leonardo de Moura <leonardo@microsoft.com>
2014-05-27 23:16:52 -07:00

463 lines
19 KiB
C++

/*
Copyright (c) 2014 Microsoft Corporation. All rights reserved.
Released under Apache 2.0 license as described in the file LICENSE.
Author: Leonardo de Moura
*/
#include "util/interrupt.h"
#include "util/lbool.h"
#include "kernel/converter.h"
#include "kernel/expr_maps.h"
#include "kernel/instantiate.h"
#include "kernel/free_vars.h"
namespace lean {
static no_delayed_justification g_no_delayed_jst;
bool converter::is_def_eq(expr const & t, expr const & s, context & c) {
return is_def_eq(t, s, c, g_no_delayed_jst);
}
/** \brief Do nothing converter */
struct dummy_converter : public converter {
virtual expr whnf(expr const & e, context &) { return e; }
virtual bool is_def_eq(expr const &, expr const &, context &, delayed_justification &) { return true; }
};
std::unique_ptr<converter> mk_dummy_converter() {
return std::unique_ptr<converter>(new dummy_converter());
}
struct default_converter : public converter {
environment m_env;
optional<module_idx> m_module_idx;
bool m_memoize;
name_set m_extra_opaque;
expr_struct_map<expr> m_whnf_core_cache;
expr_struct_map<expr> m_whnf_cache;
default_converter(environment const & env, optional<module_idx> mod_idx, bool memoize, name_set const & extra_opaque):
m_env(env), m_module_idx(mod_idx), m_memoize(memoize), m_extra_opaque(extra_opaque) {}
class extended_context : public extension_context {
default_converter & m_conv;
context & m_ctx;
public:
extended_context(default_converter & conv, context & ctx):m_conv(conv), m_ctx(ctx) {}
virtual environment const & env() const { return m_conv.m_env; }
virtual expr whnf(expr const & e) { return m_conv.whnf(e, m_ctx); }
virtual bool is_def_eq(expr const & e1, expr const & e2, delayed_justification & j) { return m_conv.is_def_eq(e1, e2, m_ctx, j); }
virtual expr infer_type(expr const & e) { return m_ctx.infer_type(e); }
virtual name mk_fresh_name() { return m_ctx.mk_fresh_name(); }
virtual void add_cnstr(constraint const & c) { m_ctx.add_cnstr(c); }
};
optional<expr> expand_macro(expr const & m, context & c) {
lean_assert(is_macro(m));
extended_context xctx(*this, c);
return macro_def(m).expand(m, xctx);
}
/** \brief Apply normalizer extensions to \c e. */
optional<expr> norm_ext(expr const & e, context & c) {
extended_context xctx(*this, c);
return m_env.norm_ext()(e, xctx);
}
/** \brief Try to apply eta-reduction to \c e. */
expr try_eta(expr const & e) {
lean_assert(is_lambda(e));
expr const & b = binding_body(e);
if (is_lambda(b)) {
expr new_b = try_eta(b);
if (is_eqp(b, new_b)) {
return e;
} else if (is_app(new_b) && is_var(app_arg(new_b), 0) && !has_free_var(app_fn(new_b), 0)) {
return lower_free_vars(app_fn(new_b), 1);
} else {
return update_binding(e, binding_domain(e), new_b);
}
} else if (is_app(b) && is_var(app_arg(b), 0) && !has_free_var(app_fn(b), 0)) {
return lower_free_vars(app_fn(b), 1);
} else {
return e;
}
}
/** \brief Weak head normal form core procedure. It does not perform delta reduction nor normalization extensions. */
expr whnf_core(expr const & e, context & c) {
check_system("whnf");
// handle easy cases
switch (e.kind()) {
case expr_kind::Var: case expr_kind::Sort: case expr_kind::Meta: case expr_kind::Local:
case expr_kind::Pi: case expr_kind::Constant:
return e;
case expr_kind::Lambda: case expr_kind::Macro: case expr_kind::Let: case expr_kind::App:
break;
}
// check cache
if (m_memoize) {
auto it = m_whnf_core_cache.find(e);
if (it != m_whnf_core_cache.end())
return it->second;
}
// do the actual work
expr r;
switch (e.kind()) {
case expr_kind::Var: case expr_kind::Sort: case expr_kind::Meta: case expr_kind::Local:
case expr_kind::Pi: case expr_kind::Constant:
lean_unreachable(); // LCOV_EXCL_LINE
case expr_kind::Lambda:
r = (m_env.eta()) ? try_eta(e) : e;
break;
case expr_kind::Macro:
if (auto m = expand_macro(e, c))
r = whnf_core(*m, c);
else
r = e;
break;
case expr_kind::Let:
r = whnf_core(instantiate(let_body(e), let_value(e)), c);
break;
case expr_kind::App: {
buffer<expr> args;
expr f0 = get_app_rev_args(e, args);
expr f = whnf_core(f0, c);
if (is_lambda(f)) {
unsigned m = 1;
unsigned num_args = args.size();
while (is_lambda(binding_body(f)) && m < num_args) {
f = binding_body(f);
m++;
}
lean_assert(m <= num_args);
r = whnf_core(mk_rev_app(instantiate(binding_body(f), m, args.data() + (num_args - m)), num_args - m, args.data()), c);
} else {
r = is_eqp(f, f0) ? e : mk_rev_app(f, args.size(), args.data());
}
break;
}}
if (m_memoize)
m_whnf_core_cache.insert(mk_pair(e, r));
return r;
}
/**
\brief Predicate for deciding whether \c d is an opaque definition or not.
Here is the basic idea:
1) Each definition has an opaque flag. This flag cannot be modified after a definition is added to the environment.
The opaque flag affects the convertability check. The idea is to minimize the number of delta-reduction steps.
We also believe it increases the modularity of Lean developments by minimizing the dependency on how things are defined.
We should view non-opaque definitions as "inline definitions" used in programming languages such as C++.
2) Whenever type checking an expression, the user can provide an additional set of definition names (m_extra_opaque) that
should be considered opaque. Note that, if \c t type checks when using an extra_opaque set S, then t also type checks
(modulo resource constraints) with the empty set. Again, the purpose of extra_opaque is to mimimize the number
of delta-reduction steps.
3) To be able to prove theorems about an opaque definition, we treat an opaque definition D in a module M as
transparent when we are type checking another definition/theorem D' also in M. This rule only applies if
D is not a theorem, nor D is in the set m_extra_opaque. To implement this feature, this class has a field
m_module_idx that is not none when this rule should be applied.
*/
bool is_opaque(declaration const & d) const {
lean_assert(d.is_definition());
if (d.is_theorem()) return true; // theorems are always opaque
if (m_extra_opaque.contains(d.get_name())) return true; // extra_opaque set overrides opaque flag
if (!d.is_opaque()) return false; // d is a transparent definition
if (m_module_idx && d.get_module_idx() == *m_module_idx) return false; // the opaque definitions in module_idx are considered transparent
return true; // d is opaque
}
/** \brief Expand \c e if it is non-opaque constant with weight >= w */
expr unfold_name_core(expr e, unsigned w) {
if (is_constant(e)) {
if (auto d = m_env.find(const_name(e))) {
if (d->is_definition() && !is_opaque(*d) && d->get_weight() >= w)
return unfold_name_core(instantiate_params(d->get_value(), d->get_params(), const_levels(e)), w);
}
}
return e;
}
/**
\brief Expand constants and application where the function is a constant.
The unfolding is only performend if the constant corresponds to
a non-opaque definition with weight >= w.
*/
expr unfold_names(expr const & e, unsigned w) {
if (is_app(e)) {
expr f0 = get_app_fn(e);
expr f = unfold_name_core(f0, w);
if (is_eqp(f, f0)) {
return e;
} else {
buffer<expr> args;
get_app_rev_args(e, args);
return mk_rev_app(f, args);
}
} else {
return unfold_name_core(e, w);
}
}
/** \brief Auxiliary method for \c is_delta */
optional<declaration> is_delta_core(expr const & e) {
if (is_constant(e)) {
if (auto d = m_env.find(const_name(e)))
if (d->is_definition() && !is_opaque(*d))
return d;
}
return none_declaration();
}
/**
\brief Return some definition \c d iff \c e is a target for delta-reduction, and the given definition is the one
to be expanded.
*/
optional<declaration> is_delta(expr const & e) { return is_delta_core(get_app_fn(e)); }
/**
\brief Weak head normal form core procedure that perform delta reduction for non-opaque constants with
weight greater than or equal to \c w.
This method is based on <tt>whnf_core(expr const &)</tt> and \c unfold_names.
\remark This method does not use normalization extensions attached in the environment.
*/
expr whnf_core(expr e, unsigned w, context & c) {
while (true) {
expr new_e = unfold_names(whnf_core(e, c), w);
if (is_eqp(e, new_e))
return e;
e = new_e;
}
}
/** \brief Put expression \c t in weak head normal form */
virtual expr whnf(expr const & e_prime, context & c) {
expr e = e_prime;
// check cache
if (m_memoize) {
auto it = m_whnf_cache.find(e);
if (it != m_whnf_cache.end())
return it->second;
}
expr t = e;
while (true) {
expr t1 = whnf_core(t, 0, c);
auto new_t = norm_ext(t1, c);
if (new_t) {
t = *new_t;
} else {
if (m_memoize)
m_whnf_cache.insert(mk_pair(e, t1));
return t1;
}
}
}
/**
\brief Given lambda/Pi expressions \c t and \c s, return true iff \c t is def eq to \c s.
t and s are definitionally equal
iff
domain(t) is definitionally equal to domain(s)
and
body(t) is definitionally equal to body(s)
*/
bool is_def_eq_binding(expr t, expr s, context & c, delayed_justification & jst) {
lean_assert(t.kind() == s.kind());
lean_assert(is_binding(t));
expr_kind k = t.kind();
buffer<expr> subst;
do {
expr var_t_type = instantiate(binding_domain(t), subst.size(), subst.data());
expr var_s_type = instantiate(binding_domain(s), subst.size(), subst.data());
if (!is_def_eq(var_t_type, var_s_type, c, jst))
return false;
subst.push_back(mk_local(c.mk_fresh_name(), binding_name(s), var_s_type));
t = binding_body(t);
s = binding_body(s);
} while (t.kind() == k && s.kind() == k);
return is_def_eq(instantiate(t, subst.size(), subst.data()), instantiate(s, subst.size(), subst.data()), c, jst);
}
/** \brief This is an auxiliary method for is_def_eq. It handles the "easy cases". */
lbool quick_is_def_eq(expr const & t, expr const & s, context & c, delayed_justification & jst) {
if (t == s)
return l_true; // t and s are structurally equal
if (is_meta(t) || is_meta(s)) {
// if t or s is a metavariable (or the application of a metavariable), then add constraint
c.add_cnstr(mk_eq_cnstr(t, s, jst.get()));
return l_true;
}
if (t.kind() == s.kind()) {
switch (t.kind()) {
case expr_kind::Lambda: case expr_kind::Pi:
return to_lbool(is_def_eq_binding(t, s, c, jst));
case expr_kind::Sort:
// t and s are Sorts
if (is_equivalent(sort_level(t), sort_level(s))) {
return l_true;
} else if (has_meta(sort_level(t)) || has_meta(sort_level(s))) {
c.add_cnstr(mk_level_cnstr(sort_level(t), sort_level(s), jst.get()));
return l_true;
} else {
return l_false;
}
case expr_kind::Meta:
lean_unreachable(); // LCOV_EXCL_LINE
case expr_kind::Var: case expr_kind::Local: case expr_kind::App:
case expr_kind::Constant: case expr_kind::Macro: case expr_kind::Let:
// We do not handle these cases in this method.
break;
}
}
return l_undef; // This is not an "easy case"
}
/**
\brief Return true if arguments of \c t are definitionally equal to arguments of \c s.
This method is used to implement an optimization in the method \c is_def_eq.
*/
bool is_def_eq_args(expr t, expr s, context & c, delayed_justification & jst) {
try {
while (is_app(t) && is_app(s)) {
if (!is_def_eq(app_arg(t), app_arg(s), c, jst))
return false;
t = app_fn(t);
s = app_fn(s);
}
return !is_app(t) && !is_app(s);
} catch (add_cnstr_exception &) {
return false;
}
}
/** \brief Return true iff t is a constant named f_name or an application of the form (f_name a_1 ... a_k) */
bool is_app_of(expr t, name const & f_name) {
t = get_app_fn(t);
return is_constant(t) && const_name(t) == f_name;
}
/** Return true iff t is definitionally equal to s. */
virtual bool is_def_eq(expr const & t, expr const & s, context & c, delayed_justification & jst) {
check_system("is_definitionally_equal");
lbool r = quick_is_def_eq(t, s, c, jst);
if (r != l_undef) return r == l_true;
// apply whnf (without using delta-reduction or normalizer extensions)
expr t_n = whnf_core(t, c);
expr s_n = whnf_core(s, c);
if (!is_eqp(t_n, t) || !is_eqp(s_n, s)) {
r = quick_is_def_eq(t_n, s_n, c, jst);
if (r != l_undef) return r == l_true;
}
// lazy delta-reduction and then normalizer extensions
while (true) {
// first, keep applying lazy delta-reduction while applicable
while (true) {
auto d_t = is_delta(t_n);
auto d_s = is_delta(s_n);
if (!d_t && !d_s) {
break;
} else if (d_t && !d_s) {
t_n = whnf_core(unfold_names(t_n, 0), c);
} else if (!d_t && d_s) {
s_n = whnf_core(unfold_names(s_n, 0), c);
} else if (d_t->get_weight() > d_s->get_weight()) {
t_n = whnf_core(unfold_names(t_n, d_s->get_weight() + 1), c);
} else if (d_t->get_weight() < d_s->get_weight()) {
s_n = whnf_core(unfold_names(s_n, d_t->get_weight() + 1), c);
} else {
lean_assert(d_t && d_s && d_t->get_weight() == d_s->get_weight());
// If t_n and s_n are both applications of the same (non-opaque) definition,
// then we try to check if their arguments are definitionally equal.
// If they are, then t_n and s_n must be definitionally equal, and we can
// skip the delta-reduction step.
// We only apply the optimization if t_n and s_n do not contain metavariables.
// In this way we don't have to backtrack constraints if the optimization fail.
if (is_app(t_n) && is_app(s_n) &&
is_eqp(*d_t, *d_s) && // same definition
!has_metavar(t_n) &&
!has_metavar(s_n) &&
!is_opaque(*d_t) && // if d_t is opaque, we don't need to try this optimization
d_t->use_conv_opt() && // the flag use_conv_opt() can be used to disable this optimization
is_def_eq_args(t_n, s_n, c, jst)) {
return true;
}
t_n = whnf_core(unfold_names(t_n, d_t->get_weight() - 1), c);
s_n = whnf_core(unfold_names(s_n, d_s->get_weight() - 1), c);
}
r = quick_is_def_eq(t_n, s_n, c, jst);
if (r != l_undef) return r == l_true;
}
// try normalizer extensions
auto new_t_n = norm_ext(t_n, c);
auto new_s_n = norm_ext(s_n, c);
if (!new_t_n && !new_s_n)
break; // t_n and s_n are in weak head normal form
if (new_t_n)
t_n = whnf_core(*new_t_n, c);
if (new_s_n)
s_n = whnf_core(*new_s_n, c);
r = quick_is_def_eq(t_n, s_n, c, jst);
if (r != l_undef) return r == l_true;
}
// At this point, t_n and s_n are in weak head normal form (modulo meta-variables and proof irrelevance)
if (is_app(t_n) && is_app(s_n)) {
expr it1 = t_n;
expr it2 = s_n;
bool ok = true;
do {
if (!is_def_eq(app_arg(it1), app_arg(it2), c, jst)) {
ok = false;
break;
}
it1 = app_fn(it1);
it2 = app_fn(it2);
} while (is_app(it1) && is_app(it2));
if (ok && is_def_eq(it1, it2, c, jst))
return true;
}
if (m_env.prop_proof_irrel()) {
// Proof irrelevance support for Prop/Bool (aka Type.{0})
expr t_type = c.infer_type(t);
if (is_prop(t_type, c) && is_def_eq(t_type, c.infer_type(s), c, jst))
return true;
}
list<name> const & cls_proof_irrel = m_env.cls_proof_irrel();
if (!is_nil(cls_proof_irrel)) {
// Proof irrelevance support for classes
expr t_type = whnf(c.infer_type(t), c);
if (std::any_of(cls_proof_irrel.begin(), cls_proof_irrel.end(),
[&](name const & cls_name) { return is_app_of(t_type, cls_name); }) &&
is_def_eq(t_type, c.infer_type(s), c, jst))
return true;
}
return false;
}
bool is_prop(expr const & e, context & c) {
return whnf(c.infer_type(e), c) == Bool;
}
};
std::unique_ptr<converter> mk_default_converter(environment const & env, optional<module_idx> mod_idx,
bool memoize, name_set const & extra_opaque) {
return std::unique_ptr<converter>(new default_converter(env, mod_idx, memoize, extra_opaque));
}
}