Implement ad hoc regex pattern rule support

An ad hoc pattern rule consists of a pattern that mimics a dependency
declaration followed by one or more recipes. For example:

exe{~'/(.*)/'}: cxx{~'/\1/'}
{{
  $cxx.path -o $path($>) $path($<[0])
}}

If a pattern matches a dependency declaration of a target, then the recipe is
used to perform the corresponding operation on this target. For example, the
following dependency declaration matches the above pattern which means the
rule's recipe will be used to update this target:

exe{hello}: cxx{hello}

While the following declarations do not match the above pattern:

exe{hello}:   c{hello}  # Type mismatch.
exe{hello}: cxx{howdy}  # Name mismatch.

On the left hand side of `:` in the pattern we can have a single target or an
ad hoc target group. The single target or the first (primary) ad hoc group
member must be a regex pattern (~). The rest of the ad hoc group members can
be patterns or substitutions (^). For example:

<exe{~'/(.*)/'} file{^'/\1.map/'}>: cxx{~'/\1/'}
{{
  $cxx.path -o $path($>[0]) "-Wl,-Map=$path($>[1])" $path($<[0])
}}

On the left hand side of `:` in the pattern we have prerequisites which can
be patterns, substitutions, or non-patterns. For example:

<exe{~'/(.*)/'} file{^'/\1.map/'}>: cxx{~'/\1/'} hxx{^'/\1/'} hxx{common}
{{
  $cxx.path -o $path($>[0]) "-Wl,-Map=$path($>[1])" $path($<[0])
}}

Substitutions on the left hand side of `:` and substitutions and non-patterns
on the right hand side are added to the dependency declaration. For example,
given the above rule and dependency declaration, the effective dependency is
going to be:

<exe{hello} file{hello.map>: cxx{hello} hxx{hello} hxx{common}

1 files changed, 442 insertions, 0 deletions

diff --git a/libbuild2/adhoc-rule-regex-pattern.cxx b/libbuild2/adhoc-rule-regex-pattern.cxx
new file mode 100644
index 0000000..4c8c1e5
--- /dev/null
+++ b/libbuild2/adhoc-rule-regex-pattern.cxx
@@ -0,0 +1,442 @@
+// file      : libbuild2/adhoc-rule-regex-pattern.cxx -*- C++ -*-
+// license   : MIT; see accompanying LICENSE file
+
+#include <libbuild2/adhoc-rule-regex-pattern.hxx>
+
+#include <libbutl/regex.mxx>
+
+#include <libbuild2/algorithm.hxx>
+
+namespace build2
+{
+  using pattern_type = name::pattern_type;
+
+  adhoc_rule_regex_pattern::
+  adhoc_rule_regex_pattern (
+    const scope& s, string rn, const target_type& tt,
+    name&& n, const location& nloc,
+    names&& ans, const location& aloc,
+    names&& pns, const location& ploc)
+      : adhoc_rule_pattern (s, move (rn), tt)
+  {
+    // Semantically, our rule pattern is one logical regular expression that
+    // spans multiple targets and prerequisites with a single back reference
+    // (\N) space.
+    //
+    // To implement this we are going to concatenate all the target and
+    // prerequisite sub-patterns separated with a character which cannot
+    // appear in the name (nor is a special regex character) but which is
+    // printable (for diagnostics). The directory separator (`/`) feels like a
+    // natural choice. We will call such a concatenated string of names a
+    // "name signature" (we also have a "type signature"; see below) and its
+    // pattern a "name signature pattern".
+    //
+    regex::flag_type flags (regex::ECMAScript);
+
+    // Append the sub-pattern to text_ returning the status of the `e` flag.
+    //
+    auto append_pattern = [this, &flags, first = true] (
+      const string& t,
+      const location& loc) mutable -> bool
+    {
+      size_t n (t.size ()), p (t.rfind (t[0]));
+
+      // Process flags.
+      //
+      bool fi (false), fe (false);
+      for (size_t i (p + 1); i != n; ++i)
+      {
+        switch (t[i])
+        {
+        case 'i': fi = true; break;
+        case 'e': fe = true; break;
+        }
+      }
+
+      // For icase we require all or none of the patterns to have it.
+      //
+      if (first)
+      {
+        if (fi)
+          flags |= regex::icase;
+      }
+      else if (((flags & regex::icase) != 0) != fi)
+        fail (loc) << "inconsistent regex 'i' flag in '" << t << "'";
+
+      if (!first)
+        text_ += '/';
+      else
+        first = false;
+
+      text_.append (t.c_str () + 1, p - 1);
+
+      return fe;
+    };
+
+    // Append an element either to targets_ or prereqs_.
+    //
+    auto append_element = [&s, &append_pattern] (
+      vector<element>& v,
+      name&& n,
+      const location& loc,
+      const target_type* tt = nullptr)
+    {
+      if (tt == nullptr)
+      {
+        tt = n.untyped () || n.type == "*"
+        ? &target::static_type
+        : s.find_target_type (n.type);
+
+        if (tt == nullptr)
+          fail (loc) << "unknown target type " << n.type;
+      }
+
+      bool e (n.pattern                                 &&
+              *n.pattern == pattern_type::regex_pattern &&
+              append_pattern (n.value, loc));
+
+      v.push_back (element {move (n), *tt, e});
+    };
+
+    // This one is always a pattern.
+    //
+    append_element (targets_, move (n), nloc, &tt);
+
+    // These are all patterns or substitutions.
+    //
+    for (name& an: ans)
+      append_element (targets_, move (an), aloc);
+
+    // These can be patterns, substitutions, or non-patterns.
+    //
+    for (name& pn: pns)
+      append_element (prereqs_, move (pn), ploc);
+
+    try
+    {
+      regex_ = regex (text_, flags);
+    }
+    catch (const regex_error& e)
+    {
+      // Print regex_error description if meaningful (no space).
+      //
+      // This may not necessarily be pointing at the actual location of the
+      // error but it should be close enough.
+      //
+      fail (nloc) << "invalid regex pattern '" << text_ << "'" << e;
+    }
+  }
+
+  bool adhoc_rule_regex_pattern::
+  match (action a, target& t, const string&, match_extra& me) const
+  {
+    tracer trace ("adhoc_rule_regex_pattern::match");
+
+    // The plan is as follows: First check the "type signature" of the target
+    // and its prerequisites (the primary target type has already been matched
+    // by the rule matching machinery). If there is a match, then concatenate
+    // their names into a "name signature" in the same way as for sub-patterns
+    // above and match that against the name signature regex pattern. If there
+    // is a match then this rule matches and the apply_*() functions should be
+    // called to process any member/prerequisite substitutions and inject them
+    // along with non-pattern prerequisites.
+    //
+    // It would be natural to perform the type match and concatenation of the
+    // names simultaneously. However, while the former should be quite cheap,
+    // the latter will most likely require dynamic allocation. To mitigate
+    // this we are going to pre-type-match the first prerequisite before
+    // concatenating any names. This should weed out most of the non-matches
+    // for sane patterns.
+    //
+    // Note also that we don't backtrack and try different combinations of the
+    // type-matching targets/prerequisites. We also ignore prerequisites
+    // marked ad hoc for type-matching.
+    //
+    auto pattern = [] (const element& e) -> bool
+    {
+      return e.name.pattern && *e.name.pattern == pattern_type::regex_pattern;
+    };
+
+    auto find_prereq = [a, &t] (const target_type& tt) -> optional<target_key>
+    {
+      // We use the standard logic that one would use in the rule::match()
+      // implementation.
+      //
+      for (prerequisite_member p: group_prerequisite_members (a, t))
+      {
+        if (include (a, t, p) == include_type::normal && p.is_a (tt))
+          return p.key ().tk;
+      }
+      return nullopt;
+    };
+
+    // Pre-type-match the first prerequisite, if any.
+    //
+    auto pe (prereqs_.end ()), pi (find_if (prereqs_.begin (), pe, pattern));
+
+    optional<target_key> pk1;
+    if (pi != pe)
+    {
+      if (!(pk1 = find_prereq (pi->type)))
+      {
+        l4 ([&]{trace << rule_name << ": no " << pi->type.name
+                      << "{} prerequisite for target " << t;});
+        return false;
+      }
+    }
+
+    // Ok, this is a potential match, start concatenating the names.
+    //
+    // Note that the regex_match_results object (which we will be passing
+    // through to apply() in the target's auxiliary data storage) contains
+    // iterators pointing to the string being matched. Which means this string
+    // must be kept around until we are done with replacing the subsitutions.
+    // In fact, we cannot even move it because this may invalidate the
+    // iterators (e.g., in case of a small string optimization). So the plan
+    // is to store the string in match_extra::buffer and regex_match_results
+    // (which we can move) in the auxiliary data storage.
+    //
+    string& ns (me.buffer);
+
+    auto append_name = [&ns, first = true] (const target_key& tk,
+                                            const element& e) mutable
+    {
+      if (!first)
+        ns += '/';
+      else
+        first = false;
+
+      ns += *tk.name;
+
+      // The same semantics as in variable_type_map::find().
+      //
+      if (tk.ext && !tk.ext->empty () &&
+          (e.match_ext ||
+           tk.type->fixed_extension == &target_extension_none ||
+           tk.type->fixed_extension == &target_extension_must))
+      {
+        ns += '.';
+        ns += *tk.ext;
+      }
+    };
+
+    // Primary target (always a pattern).
+    //
+    auto te (targets_.end ()), ti (targets_.begin ());
+    append_name (t.key (), *ti);
+
+    // Match ad hoc group members.
+    //
+    while ((ti = find_if (ti + 1, te, pattern)) != te)
+    {
+      const target* at (find_adhoc_member (t, ti->type));
+
+      if (at == nullptr)
+      {
+        l4 ([&]{trace << rule_name << ": no " << ti->type.name
+                      << "{} ad hoc target group member for target " << t;});
+        return false;
+      }
+
+      append_name (at->key (), *ti);
+    }
+
+    // Finish prerequisites.
+    //
+    if (pi != pe)
+    {
+      append_name (*pk1, *pi);
+
+      while ((pi = find_if (pi + 1, pe, pattern)) != pe)
+      {
+        optional<target_key> pk (find_prereq (pi->type));
+
+        if (!pk)
+        {
+          l4 ([&]{trace << rule_name << ": no " << pi->type.name
+                        << "{} prerequisite for target " << t;});
+          return false;
+        }
+
+        append_name (*pk, *pi);
+      }
+    }
+
+    // While it can be tempting to optimize this for patterns that don't have
+    // any substitutions (which would be most of them), keep in mind that we
+    // will also need match_results for $N variables in the recipe (or a C++
+    // rule implementation may want to access the match_results object).
+    //
+    regex_match_results mr;
+    if (!regex_match (ns, mr, regex_))
+    {
+      l4 ([&]{trace << rule_name << ": name signature '" << ns
+                    << "' does not match regex '" << text_
+                    << "' for target " << t;});
+      return false;
+    }
+
+    static_assert (sizeof (regex_match_results) <= target::data_size,
+                   "insufficient space");
+    t.data (move (mr));
+
+    return true;
+  }
+
+  static inline string
+  substitute (const target& t,
+              const regex_match_results& mr,
+              const string& s,
+              const char* what)
+  {
+    string r (butl::regex_replace_match_results (
+                mr, s.c_str () + 1, s.rfind (s[0]) - 1));
+
+    // @@ Note that while it would have been nice to print the location here,
+    //    (and also pass to search()->find_target_type()), we would need to
+    //    save location_value in each element to cover multiple declarations.
+    //
+    if (r.empty ())
+      fail << what << " substitution '" << s << "' for target " << t
+           << " results in empty name";
+
+    return r;
+  }
+
+  void adhoc_rule_regex_pattern::
+  apply_adhoc_members (action, target& t, match_extra&) const
+  {
+    const auto& mr (t.data<regex_match_results> ());
+
+    for (auto i (targets_.begin () + 1); i != targets_.end (); ++i)
+    {
+      // These are all patterns or substitutions.
+      //
+      const element& e (*i);
+
+      if (*e.name.pattern == pattern_type::regex_pattern)
+        continue;
+
+      // Similar to prerequisites below, we treat member substitutions
+      // relative to the target.
+      //
+      dir_path d;
+      if (e.name.dir.empty ())
+        d = t.dir; // Absolute and normalized.
+      else
+      {
+        if (e.name.dir.absolute ())
+          d = e.name.dir;
+        else
+          d = t.dir / e.name.dir;
+
+        d.normalize ();
+      }
+
+      // @@ TODO: currently this uses type as the ad hoc member identity.
+      //
+      add_adhoc_member (
+        t,
+        e.type,
+        move (d),
+        dir_path () /* out */,
+        substitute (t, mr, e.name.value, "ad hoc target group member"));
+    }
+  }
+
+  void adhoc_rule_regex_pattern::
+  apply_prerequisites (action a, target& t, match_extra&) const
+  {
+    const auto& mr (t.data<regex_match_results> ());
+
+    // Resolve and cache target scope lazily.
+    //
+    auto base_scope = [&t, bs = (const scope*) nullptr] () mutable
+      -> const scope&
+    {
+      if (bs == nullptr)
+        bs = &t.base_scope ();
+
+      return *bs;
+    };
+
+    // Re-create the same clean semantics as in match_prerequisite_members().
+    //
+    bool clean (a.operation () == clean_id && !t.is_a<alias> ());
+
+    auto& pts (t.prerequisite_targets[a]);
+    size_t start (pts.size ());
+
+    for (const element& e: prereqs_)
+    {
+      // While it would be nice to avoid copying here, the semantics of
+      // search() (and find_target_type() that it calls) is just too hairy to
+      // duplicate and try to optimize. It feels like most of the cases will
+      // either fall under the small string optimization or be absolute target
+      // names (e.g., imported tools).
+      //
+      // @@ Perhaps we should try to optimize the absolute target name case?
+      //
+      // Which scope should we use to resolve this prerequisite? After some
+      // meditation it feels natural to use the target's scope for patterns
+      // and the rule's scope for non-patterns.
+      //
+      name n;
+      const scope* s;
+      if (e.name.pattern)
+      {
+        if (*e.name.pattern == pattern_type::regex_pattern)
+          continue;
+
+        // Note: cannot be project-qualified.
+        //
+        n = name (e.name.dir,
+                  e.name.type,
+                  substitute (t, mr, e.name.value, "prerequisite"));
+        s = &base_scope ();
+      }
+      else
+      {
+        n = e.name;
+        s = &rule_scope;
+      }
+
+      const target& pt (search (t, move (n), *s, &e.type));
+
+      if (clean && !pt.in (*base_scope ().root_scope ()))
+        continue;
+
+      // @@ TODO: it could be handy to mark a prerequisite (e.g., a tool)
+      //    ad hoc so that it doesn't interfere with the $< list.
+      //
+      pts.push_back (prerequisite_target (&pt, false /* adhoc */));
+    }
+
+    if (start != pts.size ())
+      match_members (a, t, pts, start);
+  }
+
+  void adhoc_rule_regex_pattern::
+  dump (ostream& os) const
+  {
+    // Targets.
+    //
+    size_t tn (targets_.size ());
+
+    if (tn != 1)
+      os << '<';
+
+    for (size_t i (0); i != tn; ++i)
+      os << (i != 0 ? " " : "") << targets_[i].name;
+
+    if (tn != 1)
+      os << '>';
+
+    // Prerequisites.
+    //
+    os << ':';
+
+    for (size_t i (0); i != prereqs_.size (); ++i)
+      os << ' ' << prereqs_[i].name;
+  }
+}