From 886868dc67e069734b44d81d9f56d48a0a47538e Mon Sep 17 00:00:00 2001 From: Karen Arutyunov Date: Tue, 23 Aug 2022 23:16:40 +0300 Subject: Split pkg-build.cxx --- bpkg/pkg-build-collect.cxx | 6411 ++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 6411 insertions(+) create mode 100644 bpkg/pkg-build-collect.cxx (limited to 'bpkg/pkg-build-collect.cxx') diff --git a/bpkg/pkg-build-collect.cxx b/bpkg/pkg-build-collect.cxx new file mode 100644 index 0000000..e7f5c80 --- /dev/null +++ b/bpkg/pkg-build-collect.cxx @@ -0,0 +1,6411 @@ +// file : bpkg/pkg-build-collect.cxx -*- C++ -*- +// license : MIT; see accompanying LICENSE file + +#include + +#include +#include +#include // numeric_limits +#include // cout +#include // ref() +#include + +#include +#include +#include +#include +#include + +#include + +#include +#include +#include +#include + +using namespace std; + +namespace bpkg +{ + // build_package + // + bool build_package:: + user_selection () const + { + return required_by.find (package_key {db.get ().main_database (), ""}) != + required_by.end (); + } + + bool build_package:: + user_selection (const vector& hold_pkgs) const + { + return find_if (hold_pkgs.begin (), hold_pkgs.end (), + [this] (const build_package& p) + { + return p.db == db && p.name () == name (); + }) != hold_pkgs.end (); + } + + string build_package:: + available_name_version_db () const + { + const string& s (db.get ().string); + return !s.empty () + ? available_name_version () + ' ' + s + : available_name_version (); + } + + bool build_package:: + recollect_recursively (const repointed_dependents& rpt_depts) const + { + assert (action && + *action == build_package::build && + available != nullptr && + selected != nullptr && + selected->state == package_state::configured && + selected->substate != package_substate::system); + + // Note that if the skeleton is present then the package is either being + // already collected or its configuration has been negotiated between the + // dependents. + // + return !system && + (dependencies || + selected->version != available_version () || + ((!config_vars.empty () || skeleton) && + has_buildfile_clause (available->dependencies)) || + rpt_depts.find (package_key (db, name ())) != rpt_depts.end ()); + } + + bool build_package:: + reconfigure () const + { + assert (action && *action != drop); + + return selected != nullptr && + selected->state == package_state::configured && + ((flags & adjust_reconfigure) != 0 || + (*action == build && + (selected->system () != system || + selected->version != available_version () || + (!system && (!config_vars.empty () || disfigure))))); + } + + bool build_package:: + configure_only () const + { + assert (action); + + return configure_only_ || + (*action == build && (flags & (build_repoint | build_reevaluate)) != 0); + } + + const version& build_package:: + available_version () const + { + // This should have been diagnosed before creating build_package object. + // + assert (available != nullptr && + (system + ? available->system_version (db) != nullptr + : !available->stub ())); + + return system ? *available->system_version (db) : available->version; + } + + bool build_package:: + external (dir_path* d) const + { + assert (action); + + if (*action == build_package::drop) + return false; + + // If adjustment or orphan, then new and old are the same. + // + if (available == nullptr || available->locations.empty ()) + { + assert (selected != nullptr); + + if (selected->external ()) + { + assert (selected->src_root); + + if (d != nullptr) + *d = *selected->src_root; + + return true; + } + } + else + { + const package_location& pl (available->locations[0]); + + if (pl.repository_fragment.object_id () == "") // Special root? + { + if (!exists (pl.location)) // Directory case? + { + if (d != nullptr) + *d = normalize (path_cast (pl.location), "package"); + + return true; + } + } + else + { + // See if the package comes from the directory-based repository, and + // so is external. + // + // Note that such repository fragments are always preferred over + // others (see below). + // + for (const package_location& pl: available->locations) + { + const repository_location& rl ( + pl.repository_fragment.load ()->location); + + if (rl.directory_based ()) + { + // Note that the repository location path is always absolute for + // the directory-based repositories but the package location may + // potentially not be normalized. Thus, we normalize the resulting + // path, if requested. + // + if (d != nullptr) + *d = normalize (path_cast (rl.path () / pl.location), + "package"); + + return true; + } + } + } + } + + return false; + } + + void build_package:: + merge (build_package&& p) + { + // We don't merge objects from different configurations. + // + assert (db == p.db); + + // We don't merge into pre-entered objects, and from/into drops. + // + assert (action && *action != drop && (!p.action || *p.action != drop)); + + // We never merge two repointed dependent reconfigurations. + // + assert ((flags & build_repoint) == 0 || (p.flags & build_repoint) == 0); + + // We never merge two existing dependent re-evaluations. + // + assert ((flags & build_reevaluate) == 0 || + (p.flags & build_reevaluate) == 0); + + // Copy the user-specified options/variables. + // + if (p.user_selection ()) + { + // We don't allow a package specified on the command line multiple times + // to have different sets of options/variables. Given that, it's + // tempting to assert that the options/variables don't change if we + // merge into a user selection. That's, however, not the case due to the + // iterative plan refinement implementation details (--checkout-* + // options and variables are only saved into the pre-entered + // dependencies, etc.). + // + // Note that configuration can only be specified for packages on the + // command line and such packages get collected/pre-entered early, + // before any prerequisites get collected. Thus, it doesn't seem + // possible that a package configuration/options may change after we + // have created the package skeleton. + // + // Also note that if it wouldn't be true, we would potentially need to + // re-collect the package prerequisites, since configuration change + // could affect the enable condition evaluation and, as a result, the + // dependency alternative choice. + // + assert (!skeleton || + ((p.config_vars.empty () || p.config_vars == config_vars) && + p.disfigure == disfigure)); + + if (p.keep_out) + keep_out = p.keep_out; + + if (p.disfigure) + disfigure = p.disfigure; + + if (p.configure_only_) + configure_only_ = p.configure_only_; + + if (p.checkout_root) + checkout_root = move (p.checkout_root); + + if (p.checkout_purge) + checkout_purge = p.checkout_purge; + + if (!p.config_vars.empty ()) + config_vars = move (p.config_vars); + + // Propagate the user-selection tag. + // + required_by.emplace (db.get ().main_database (), ""); + } + + // Copy the required-by package names only if semantics matches. + // + if (p.required_by_dependents == required_by_dependents) + required_by.insert (p.required_by.begin (), p.required_by.end ()); + + // Copy constraints. + // + // Note that we may duplicate them, but this is harmless. + // + constraints.insert (constraints.end (), + make_move_iterator (p.constraints.begin ()), + make_move_iterator (p.constraints.end ())); + + // Copy hold_* flags if they are "stronger". + // + if (!hold_package || (p.hold_package && *p.hold_package > *hold_package)) + hold_package = p.hold_package; + + if (!hold_version || (p.hold_version && *p.hold_version > *hold_version)) + hold_version = p.hold_version; + + // Copy state flags. + // + flags |= p.flags; + + // Upgrade dependent repointments and re-evaluations to the full builds. + // + if (*action == build) + flags &= ~(build_repoint | build_reevaluate); + + // Note that we don't copy the build_package::system flag. If it was + // set from the command line ("strong system") then we will also have + // the '==' constraint which means that this build_package object will + // never be replaced. + // + // For other cases ("weak system") we don't want to copy system over in + // order not prevent, for example, system to non-system upgrade. + } + + package_skeleton& build_package:: + init_skeleton (const common_options& options, + const shared_ptr& override) + { + shared_ptr ap (override != nullptr + ? override + : available); + + assert (!skeleton && ap != nullptr); + + package_key pk (db, ap->id.name); + + if (system) + { + // Keep the available package if its version is "close enough" to the + // system package version. For now we will require the exact match + // but in the future we could relax this (e.g., allow the user to + // specify something like libfoo/^1.2.0 or some such). + // + const version* v (!ap->stub () ? ap->system_version (db) : nullptr); + + if (v == nullptr || *v != ap->version) + ap = nullptr; + } + + optional src_root, out_root; + + if (ap != nullptr) + { + src_root = external_dir (); + out_root = (src_root && !disfigure + ? dir_path (db.get ().config) /= name ().string () + : optional ()); + } + + skeleton = package_skeleton ( + options, + move (pk), + system, + move (ap), + config_vars, // @@ Maybe make optional and move? + disfigure, + (selected != nullptr ? &selected->config_variables : nullptr), + move (src_root), + move (out_root)); + + return *skeleton; + } + + // replaced_versions + // + void replaced_versions:: + cancel_bogus (tracer& trace, bool scratch) + { + bool bogus (false); + for (auto i (begin ()); i != end (); ) + { + const replaced_version& v (i->second); + + if (!v.replaced) + { + bogus = true; + + l5 ([&]{trace << "erase bogus version replacement " << i->first;}); + + i = erase (i); + } + else + ++i; + } + + if (bogus && scratch) + { + l5 ([&]{trace << "bogus version replacement erased, throwing";}); + throw cancel_replacement (); + } + } + + // postponed_configuration + // + postponed_configuration::dependency* + postponed_configuration::dependent_info:: + find_dependency (pair pos) + { + auto i (find_if (dependencies.begin (), dependencies.end (), + [&pos] (const dependency& d) + { + return d.position == pos; + })); + return i != dependencies.end () ? &*i : nullptr; + } + + void postponed_configuration::dependent_info:: + add (dependency&& dep) + { + if (dependency* d = find_dependency (dep.position)) + { + // Feels like we can accumulate dependencies into an existing + // position only for an existing dependent. + // + assert (existing); + + for (package_key& p: dep) + { + // Add the dependency unless it's already there. + // + if (find (d->begin (), d->end (), p) == d->end ()) + d->push_back (move (p)); + } + + // Set the has_alternative flag for an existing dependent. Note that + // it shouldn't change if already set. + // + if (dep.has_alternative) + { + if (!d->has_alternative) + d->has_alternative = *dep.has_alternative; + else + assert (*d->has_alternative == *dep.has_alternative); + } + } + else + dependencies.push_back (move (dep)); + } + + void postponed_configuration:: + add (package_key&& dependent, + bool existing, + pair position, + packages&& deps, + optional has_alternative) + { + assert (position.first != 0 && position.second != 0); + + add_dependencies (deps); // Don't move from since will be used later. + + auto i (dependents.find (dependent)); + + if (i != dependents.end ()) + { + dependent_info& ddi (i->second); + + ddi.add (dependency (position, move (deps), has_alternative)); + + // Conceptually, on the first glance, we can only move from existing to + // non-existing (e.g., due to a upgrade/downgrade later) and that case + // is handled via the version replacement rollback. However, after + // re-evaluation the existing dependent is handled similar to the new + // dependent and we can potentially up-negotiate the dependency + // configuration for it. + // + assert (ddi.existing || !existing); + } + else + { + small_vector ds ({ + dependency (position, move (deps), has_alternative)}); + + dependents.emplace (move (dependent), + dependent_info {existing, move (ds)}); + } + } + + bool postponed_configuration:: + contains_dependency (const packages& ds) const + { + for (const package_key& d: ds) + { + if (contains_dependency (d)) + return true; + } + + return false; + } + + bool postponed_configuration:: + contains_dependency (const postponed_configuration& c) const + { + for (const auto& d: c.dependencies) + { + if (contains_dependency (d)) + return true; + } + + return false; + } + + const pair* postponed_configuration:: + existing_dependent_position (const package_key& p) const + { + const pair* r (nullptr); + + auto i (dependents.find (p)); + if (i != dependents.end () && i->second.existing) + { + for (const dependency& d: i->second.dependencies) + { + if (r == nullptr || d.position < *r) + r = &d.position; + } + + assert (r != nullptr); + } + + return r; + } + + void postponed_configuration:: + merge (postponed_configuration&& c) + { + assert (c.id != id); // Can't merge to itself. + + merged_ids.push_back (c.id); + + // Merge dependents. + // + for (auto& d: c.dependents) + { + auto i (dependents.find (d.first)); + + if (i != dependents.end ()) + { + dependent_info& ddi (i->second); // Destination dependent info. + dependent_info& sdi (d.second); // Source dependent info. + + for (dependency& sd: sdi.dependencies) + ddi.add (move (sd)); + + // As in add() above. + // + assert (ddi.existing || !sdi.existing); + } + else + dependents.emplace (d.first, move (d.second)); + } + + // Merge dependencies. + // + add_dependencies (move (c.dependencies)); + + // Pick the depth of the outermost negotiated configuration (minimum + // non-zero depth) between the two. + // + if (depth != 0) + { + if (c.depth != 0 && depth > c.depth) + depth = c.depth; + } + else + depth = c.depth; + } + + void postponed_configuration:: + set_shadow_cluster (postponed_configuration&& c) + { + shadow_cluster.clear (); + + for (auto& dt: c.dependents) + { + positions ps; + for (auto& d: dt.second.dependencies) + ps.push_back (d.position); + + shadow_cluster.emplace (dt.first, move (ps)); + } + } + + bool postponed_configuration:: + contains_in_shadow_cluster (package_key dependent, + pair pos) const + { + auto i (shadow_cluster.find (dependent)); + + if (i != shadow_cluster.end ()) + { + const positions& ps (i->second); + return find (ps.begin (), ps.end (), pos) != ps.end (); + } + else + return false; + } + + std::string postponed_configuration:: + string () const + { + std::string r; + + for (const auto& d: dependents) + { + r += r.empty () ? '{' : ' '; + r += d.first.string (); + + if (d.second.existing) + r += '^'; + } + + if (r.empty ()) + r += '{'; + + r += " |"; + + for (const package_key& d: dependencies) + { + r += ' '; + r += d.string (); + r += "->{"; + + bool first (true); + for (const auto& dt: dependents) + { + for (const dependency& dp: dt.second.dependencies) + { + if (find (dp.begin (), dp.end (), d) != dp.end ()) + { + if (!first) + r += ' '; + else + first = false; + + r += dt.first.string (); + r += '/'; + r += to_string (dp.position.first); + r += ','; + r += to_string (dp.position.second); + } + } + } + + r += '}'; + } + + r += '}'; + + if (negotiated) + r += *negotiated ? '!' : '?'; + + return r; + } + + void postponed_configuration:: + add_dependencies (packages&& deps) + { + for (auto& d: deps) + { + if (find (dependencies.begin (), dependencies.end (), d) == + dependencies.end ()) + dependencies.push_back (move (d)); + } + } + + void postponed_configuration:: + add_dependencies (const packages& deps) + { + for (const auto& d: deps) + { + if (find (dependencies.begin (), dependencies.end (), d) == + dependencies.end ()) + dependencies.push_back (d); + } + } + + pair> postponed_configurations:: + add (package_key dependent, + bool existing, + pair position, + postponed_configuration::packages dependencies, + optional has_alternative) + { + tracer trace ("postponed_configurations::add"); + + assert (!dependencies.empty ()); + + // The plan is to first go through the existing clusters and check if any + // of them contain this dependent/dependencies in their shadow + // clusters. If such a cluster is found, then force-add them to + // it. Otherwise, if any dependency-intersecting clusters are present, + // then add the specified dependent/dependencies to the one with the + // minimum non-zero depth, if any, and to the first one otherwise. + // Otherwise, add the new cluster. Afterwards, merge into the resulting + // cluster other dependency-intersecting clusters. Note that in case of + // shadow, this should normally not happen because such a cluster should + // have been either pre-merged or its dependents should be in the + // cluster. But it feels like it may still happen if things change, in + // which case we will throw again (admittedly a bit fuzzy). + // + iterator ri; + bool rb (true); + + // Note that if a single dependency is added, then it can only belong to a + // single existing cluster and so no clusters merge can happen, unless we + // are force-adding. In the later case we can only merge once for a single + // dependency. + // + // Let's optimize for the common case based on these facts. + // + bool single (dependencies.size () == 1); + + // Merge dependency-intersecting clusters in the specified range into the + // resulting cluster and reset change rb to false if any of the merged in + // clusters is non-negotiated or is being negotiated. + // + // The iterator arguments refer to entries before and after the range + // endpoints, respectively. + // + auto merge = [&trace, &ri, &rb, single, this] (iterator i, + iterator e, + bool shadow_based) + { + postponed_configuration& rc (*ri); + + iterator j (i); + + // Merge the intersecting configurations. + // + bool merged (false); + for (++i; i != e; ++i) + { + postponed_configuration& c (*i); + + if (c.contains_dependency (rc)) + { + if (!c.negotiated || !*c.negotiated) + rb = false; + + l5 ([&]{trace << "merge " << c << " into " << rc;}); + + assert (!shadow_based || (c.negotiated && *c.negotiated)); + + rc.merge (move (c)); + c.dependencies.clear (); // Mark as merged from (see above). + + merged = true; + + if (single) + break; + } + } + + // Erase configurations which we have merged from. + // + if (merged) + { + i = j; + + for (++i; i != e; ) + { + if (!i->dependencies.empty ()) + { + ++i; + ++j; + } + else + i = erase_after (j); + } + } + }; + + auto trace_add = [&trace, &dependent, existing, position, &dependencies] + (const postponed_configuration& c, bool shadow) + { + if (verb >= 5) + { + diag_record dr (trace); + dr << "add {" << dependent; + + if (existing) + dr << '^'; + + dr << ' ' << position.first << ',' << position.second << ':'; + + for (const auto& d: dependencies) + dr << ' ' << d; + + dr << "} to " << c; + + if (shadow) + dr << " (shadow cluster-based)"; + } + }; + + // Try to add based on the shadow cluster. + // + { + auto i (begin ()); + for (; i != end (); ++i) + { + postponed_configuration& c (*i); + + if (c.contains_in_shadow_cluster (dependent, position)) + { + trace_add (c, true /* shadow */); + + c.add (move (dependent), + existing, + position, + move (dependencies), + has_alternative); + + break; + } + } + + if (i != end ()) + { + // Note that the cluster with a shadow cluster is by definition + // either being negotiated or has been negotiated. Actually, there + // is also a special case when we didn't negotiate the configuration + // yet and are in the process of re-evaluating existing dependents. + // Note though, that in this case we have already got the try/catch + // frame corresponding to the cluster negotiation (see + // collect_build_postponed() for details). + // + assert (i->depth != 0); + + ri = i; + + merge (before_begin (), ri, true /* shadow_based */); + merge (ri, end (), true /* shadow_based */); + + return make_pair (ref (*ri), optional ()); + } + } + + // Find the cluster to add the dependent/dependencies to. + // + optional depth; + + auto j (before_begin ()); // Precedes iterator i. + for (auto i (begin ()); i != end (); ++i, ++j) + { + postponed_configuration& c (*i); + + if (c.contains_dependency (dependencies) && + (!depth || (c.depth != 0 && (*depth == 0 || *depth > c.depth)))) + { + ri = i; + depth = c.depth; + } + } + + if (!depth) // No intersecting cluster? + { + // New cluster. Insert after the last element. + // + ri = insert_after (j, + postponed_configuration ( + next_id_++, + move (dependent), + existing, + position, + move (dependencies), + has_alternative)); + + l5 ([&]{trace << "create " << *ri;}); + } + else + { + // Add the dependent/dependencies into an existing cluster. + // + postponed_configuration& c (*ri); + + trace_add (c, false /* shadow */); + + c.add (move (dependent), + existing, + position, + move (dependencies), + has_alternative); + + // Try to merge other clusters into this cluster. + // + merge (before_begin (), ri, false /* shadow_based */); + merge (ri, end (), false /* shadow_based */); + } + + return make_pair (ref (*ri), optional (rb)); + } + + void postponed_configurations:: + add (package_key dependent, + pair position, + package_key dependency) + { + tracer trace ("postponed_configurations::add"); + + // Add the new cluster to the end of the list which we can only find by + // traversing the list. While at it, make sure that the dependency doesn't + // belong to any existing cluster. + // + auto i (before_begin ()); // Insert after this element. + + for (auto j (begin ()); j != end (); ++i, ++j) + assert (!j->contains_dependency (dependency)); + + i = insert_after (i, + postponed_configuration (next_id_++, + move (dependent), + position, + move (dependency))); + + l5 ([&]{trace << "create " << *i;}); + } + + postponed_configuration* postponed_configurations:: + find (size_t id) + { + for (postponed_configuration& cfg: *this) + { + if (cfg.id == id) + return &cfg; + } + + return nullptr; + } + + const postponed_configuration* postponed_configurations:: + find_dependency (const package_key& d) const + { + for (const postponed_configuration& cfg: *this) + { + if (cfg.contains_dependency (d)) + return &cfg; + } + + return nullptr; + } + + bool postponed_configurations:: + negotiated () const + { + for (const postponed_configuration& cfg: *this) + { + if (!cfg.negotiated || !*cfg.negotiated) + return false; + } + + return true; + } + + postponed_configuration& postponed_configurations:: + operator[] (size_t index) + { + auto i (begin ()); + for (size_t j (0); j != index; ++j, ++i) assert (i != end ()); + + assert (i != end ()); + return *i; + } + + size_t postponed_configurations:: + size () const + { + size_t r (0); + for (auto i (begin ()); i != end (); ++i, ++r) ; + return r; + } + + // build_packages + // + bool build_packages::package_ref:: + operator== (const package_ref& v) + { + return name == v.name && db == v.db; + } + + build_packages:: + build_packages (const build_packages& v) + : build_package_list () + { + // Copy the map. + // + for (const auto& p: v.map_) + map_.emplace (p.first, data_type {end (), p.second.package}); + + // Copy the list. + // + for (const auto& p: v) + { + auto i (map_.find (p.get ().db, p.get ().name ())); + assert (i != map_.end ()); + i->second.position = insert (end (), i->second.package); + } + } + + build_packages& build_packages:: + operator= (build_packages&& v) + { + clear (); + + // Move the map. + // + // Similar to what we do in the copy-constructor, but here we also need to + // restore the database reference and the package shared pointers in the + // source entry after the move. This way we can obtain the source packages + // databases and names later while copying the list. + // + for (auto& p: v.map_) + { + build_package& bp (p.second.package); + + database& db (bp.db); + shared_ptr sp (bp.selected); + shared_ptr ap (bp.available); + + map_.emplace (p.first, data_type {end (), move (bp)}); + + bp.db = db; + bp.selected = move (sp); + bp.available = move (ap); + } + + // Copy the list. + // + for (const auto& p: v) + { + auto i (map_.find (p.get ().db, p.get ().name ())); + assert (i != map_.end ()); + i->second.position = insert (end (), i->second.package); + } + + return *this; + } + + void build_packages:: + enter (package_name name, build_package pkg) + { + assert (!pkg.action); + + database& db (pkg.db); // Save before the move() call. + auto p (map_.emplace (package_key {db, move (name)}, + data_type {end (), move (pkg)})); + + assert (p.second); + } + + build_package* build_packages:: + collect_build (const pkg_build_options& options, + build_package pkg, + const function& fdb, + const repointed_dependents& rpt_depts, + const function& apc, + bool initial_collection, + replaced_versions& replaced_vers, + postponed_configurations& postponed_cfgs, + build_package_refs* dep_chain, + postponed_packages* postponed_repo, + postponed_packages* postponed_alts, + postponed_dependencies* postponed_deps, + postponed_positions* postponed_poss, + unacceptable_alternatives* unacceptable_alts, + const function& vpb) + { + using std::swap; // ...and not list::swap(). + + tracer trace ("collect_build"); + + // See the above notes. + // + bool recursive (dep_chain != nullptr); + assert ((postponed_repo != nullptr) == recursive && + (postponed_alts != nullptr) == recursive && + (postponed_deps != nullptr) == recursive && + (postponed_poss != nullptr) == recursive && + (unacceptable_alts != nullptr) == recursive); + + // Only builds are allowed here. + // + assert (pkg.action && *pkg.action == build_package::build && + pkg.available != nullptr); + + package_key pk (pkg.db, pkg.available->id.name); + + // Apply the version replacement, if requested, and indicate that it was + // applied. + // + auto vi (replaced_vers.find (pk)); + + if (vi != replaced_vers.end () && !vi->second.replaced) + { + l5 ([&]{trace << "apply version replacement for " + << pkg.available_name_version_db ();}); + + replaced_version& v (vi->second); + + v.replaced = true; + + if (v.available != nullptr) + { + pkg.available = v.available; + pkg.repository_fragment = v.repository_fragment; + pkg.system = v.system; + + l5 ([&]{trace << "replacement: " + << pkg.available_name_version_db ();}); + } + else + { + l5 ([&]{trace << "replacement: drop";}); + + assert (pkg.selected != nullptr); + + collect_drop (options, pkg.db, pkg.selected, replaced_vers); + return nullptr; + } + } + + // Add the version replacement entry, call the verification function if + // specified, and throw replace_version. + // + auto replace_ver = [&pk, &vpb, &vi, &replaced_vers] + (const build_package& p) + { + replaced_version rv (p.available, p.repository_fragment, p.system); + + if (vi != replaced_vers.end ()) + vi->second = move (rv); + else + replaced_vers.emplace (move (pk), move (rv)); + + if (vpb) + vpb (p, true /* scratch */); + + throw replace_version (); + }; + + auto i (map_.find (pk)); + + // If we already have an entry for this package name, then we have to + // pick one over the other. + // + // If the existing entry is a drop, then we override it. If the existing + // entry is a pre-entered or is non-build one, then we merge it into the + // new build entry. Otherwise (both are builds), we pick one and merge + // the other into it. + // + if (i != map_.end ()) + { + build_package& bp (i->second.package); + + // Note that we used to think that the scenario when the build could + // replace drop could never happen since we would start collecting + // from scratch. This has changed when we introduced replaced_versions + // for collecting drops. + // + if (bp.action && *bp.action == build_package::drop) // Drop. + { + bp = move (pkg); + } + else if (!bp.action || *bp.action != build_package::build) // Non-build. + { + pkg.merge (move (bp)); + bp = move (pkg); + } + else // Build. + { + // At the end we want p1 to point to the object that we keep + // and p2 to the object that we merge from. + // + build_package* p1 (&bp); + build_package* p2 (&pkg); + + // Pick with the following preference order: user selection over + // implicit one, source package over a system one, newer version + // over an older one. So get the preferred into p1 and the other + // into p2. + // + { + int us (p1->user_selection () - p2->user_selection ()); + int sf (p1->system - p2->system); + + if (us < 0 || + (us == 0 && sf > 0) || + (us == 0 && + sf == 0 && + p2->available_version () > p1->available_version ())) + swap (p1, p2); + } + + // If the versions differ, pick the satisfactory one and if both are + // satisfactory, then keep the preferred. + // + if (p1->available_version () != p2->available_version ()) + { + using constraint_type = build_package::constraint_type; + + // See if pv's version satisfies pc's constraints. Return the + // pointer to the unsatisfied constraint or NULL if all are + // satisfied. + // + auto test = [] (build_package* pv, + build_package* pc) -> const constraint_type* + { + for (const constraint_type& c: pc->constraints) + { + if (!satisfies (pv->available_version (), c.value)) + return &c; + } + + return nullptr; + }; + + // First see if p1 satisfies p2's constraints. + // + if (auto c2 = test (p1, p2)) + { + // If not, try the other way around. + // + if (auto c1 = test (p2, p1)) + { + const package_name& n (i->first.name); + const string& d1 (c1->dependent); + const string& d2 (c2->dependent); + + fail << "unable to satisfy constraints on package " << n << + info << d1 << c1->db << " depends on (" << n << " " + << c1->value << ")" << + info << d2 << c2->db << " depends on (" << n << " " + << c2->value << ")" << + info << "available " << p1->available_name_version () << + info << "available " << p2->available_name_version () << + info << "explicitly specify " << n << " version to manually " + << "satisfy both constraints"; + } + else + swap (p1, p2); + } + + l4 ([&]{trace << "pick " << p1->available_name_version_db () + << " over " << p2->available_name_version_db ();}); + } + + // See if we are replacing the object. If not, then we don't need to + // collect its prerequisites since that should have already been + // done. Remember, p1 points to the object we want to keep. + // + bool replace (p1 != &i->second.package); + + if (replace) + { + swap (*p1, *p2); + swap (p1, p2); // Setup for merge below. + } + + p1->merge (move (*p2)); + + if (replace) + { + if (p1->available_version () != p2->available_version () || + p1->system != p2->system) + { + // See if in-place replacement is possible (no dependencies, etc) + // and set scratch to false if that's the case. + // + // Firstly, such a package should not participate in any + // configuration negotiation. + // + // Other than that, it looks like the only optimization we can do + // easily is if the package has no dependencies (and thus cannot + // impose any constraints). Anything more advanced would require + // analyzing our dependencies (which we currently cannot easily + // get) and (1) either dropping the dependency build_package + // altogether if we are the only dependent (so that it doesn't + // influence any subsequent dependent) or (2) making sure our + // constraint is a sub-constraint of any other constraint and + // removing it from the dependency build_package. Maybe/later. + // + // NOTE: remember to update collect_drop() if changing anything + // here. + // + bool scratch (true); + + // While checking if the package has any dependencies skip the + // toolchain build-time dependencies since they should be quite + // common. + // + if (!has_dependencies (options, p2->available->dependencies)) + scratch = false; + + l5 ([&]{trace << p2->available_name_version_db () + << " package version needs to be replaced " + << (!scratch ? "in-place " : "") << "with " + << p1->available_name_version_db ();}); + + if (scratch) + replace_ver (*p1); + } + else + { + // It doesn't seem possible that replacing the build object + // without changing the package version may result in changing the + // package configuration since the configuration always gets into + // the initial package build entry (potentially pre-entered, + // etc). If it wouldn't be true then we would also need to add the + // replacement version entry and re-collect from scratch. + } + } + else + return nullptr; + } + } + else + { + // Treat the replacement of the existing dependent that is participating + // in the configuration negotiation also as a version replacement. This + // way we will not be treating the dependent as an existing on the + // re-collection (see query_existing_dependents() for details). + // + // Note: an existing dependent may not be configured as system. + // + if (pkg.selected != nullptr && + (pkg.selected->version != pkg.available_version () || + pkg.system)) + { + + for (const postponed_configuration& cfg: postponed_cfgs) + { + auto i (cfg.dependents.find (pk)); + + if (i != cfg.dependents.end () && i->second.existing) + replace_ver (pkg); + } + } + + // This is the first time we are adding this package name to the map. + // + l4 ([&]{trace << "add " << pkg.available_name_version_db ();}); + + i = map_.emplace (move (pk), data_type {end (), move (pkg)}).first; + } + + build_package& p (i->second.package); + + if (vpb) + vpb (p, false /* scratch */); + + // Recursively collect build prerequisites, if requested. + // + // Note that detecting dependency cycles during the satisfaction phase + // would be premature since they may not be present in the final package + // list. Instead we check for them during the ordering phase. + // + // The question, of course, is whether we can still end up with an + // infinite recursion here? Note that for an existing map entry we only + // recurse after the entry replacement. The infinite recursion would mean + // that we may replace a package in the map with the same version multiple + // times: + // + // ... p1 -> p2 -> ... p1 + // + // Every replacement increases the entry version and/or tightens the + // constraints the next replacement will need to satisfy. It feels + // impossible that a package version can "return" into the map being + // replaced once. So let's wait until some real use case proves this + // reasoning wrong. + // + if (recursive) + collect_build_prerequisites (options, + p, + fdb, + rpt_depts, + apc, + initial_collection, + replaced_vers, + *dep_chain, + postponed_repo, + postponed_alts, + 0 /* max_alt_index */, + *postponed_deps, + postponed_cfgs, + *postponed_poss, + *unacceptable_alts); + + return &p; + } + + void build_packages:: + collect_build_prerequisites (const pkg_build_options& options, + build_package& pkg, + const function& fdb, + const repointed_dependents& rpt_depts, + const function& apc, + bool initial_collection, + replaced_versions& replaced_vers, + build_package_refs& dep_chain, + postponed_packages* postponed_repo, + postponed_packages* postponed_alts, + size_t max_alt_index, + postponed_dependencies& postponed_deps, + postponed_configurations& postponed_cfgs, + postponed_positions& postponed_poss, + unacceptable_alternatives& unacceptable_alts, + pair reeval_pos) + { + // NOTE: don't forget to update collect_build_postponed() if changing + // anything in this function. + // + tracer trace ("collect_build_prerequisites"); + + assert (pkg.action && *pkg.action == build_package::build); + + const package_name& nm (pkg.name ()); + database& pdb (pkg.db); + package_key pk (pdb, nm); + + bool reeval (reeval_pos.first != 0); + + // The being re-evaluated dependent cannot be recursively collected yet. + // Also, we don't expect it being configured as system. + // + // Note that the configured package can still be re-evaluated after + // collect_build_prerequisites() has been called but didn't end up with + // the recursive collection. + // + assert (!reeval || + ((!pkg.recursive_collection || + !pkg.recollect_recursively (rpt_depts)) && + !pkg.skeleton && !pkg.system)); + + // If this package is not being re-evaluated, is not yet collected + // recursively, needs to be reconfigured, and is not yet postponed, then + // check if it is a dependency of any dependent with configuration clause + // and postpone the collection if that's the case. + // + // The reason why we don't need to do this for the re-evaluated case is as + // follows: this logic is used for an existing dependent that is not + // otherwise built (e.g., reconfigured) which means its externally- + // imposed configuration (user, dependents) is not being changed. + // + if (!reeval && + !pkg.recursive_collection && + pkg.reconfigure () && + postponed_cfgs.find_dependency (pk) == nullptr) + { + // If the dependent is being built, then check if it was re-evaluated to + // the position greater than the dependency position. Return true if + // that's the case, so this package is added to the resulting list and + // we can handle this situation below. + // + // Note that we rely on "small function object" optimization here. + // + const function verify ( + [&postponed_cfgs] + (const package_key& pk, pair pos) + { + for (const postponed_configuration& cfg: postponed_cfgs) + { + if (cfg.negotiated) + { + if (const pair* p = + cfg.existing_dependent_position (pk)) + { + if (p->first > pos.first) + return true; + } + } + } + + return false; + }); + + // Note that there can be multiple existing dependents for a dependency. + // Strictly speaking, we only need to add the first one with the + // assumption that the remaining dependents will also be considered + // comes the time for the negotiation. Let's, however, process all of + // them to detect the potential "re-evaluation on the greater dependency + // index" situation earlier. And, generally, have as much information as + // possible up front. + // + vector eds ( + query_existing_dependents (trace, + pk.db, + pk.name, + replaced_vers, + rpt_depts, + verify)); + + if (!eds.empty ()) + { + for (existing_dependent& ed: eds) + { + package_key dpk (ed.db, ed.selected->name); + size_t& di (ed.dependency_position.first); + + const build_package* bp (&pkg); + + // Check if this dependent needs to be re-evaluated to an earlier + // dependency position and, if that's the case, create the + // configuration cluster with this dependency instead. + // + // Note that if the replace flag is false, we proceed normally with + // the assumption that the dependency referred by the entry will be + // collected later and its configuration cluster will be created + // normally and will be negotiated earlier than the cluster being + // created for the current dependency (see collect_build_postponed() + // for details). + // + { + auto pi (postponed_poss.find (dpk)); + + if (pi != postponed_poss.end () && pi->second.first < di) + { + // If requested, override the first encountered non-replace + // position to replace. See collect_build_postponed () for + // details. + // + if (!pi->second.replace && postponed_poss.replace) + { + pi->second.replace = true; + postponed_poss.replace = false; + } + + if (pi->second.replace) + { + // Overwrite the existing dependent dependency information and + // fall through to proceed as for the normal case. + // + bp = replace_existing_dependent_dependency ( + trace, + options, + ed, // Note: modified. + pi->second, + fdb, + rpt_depts, + apc, + initial_collection, + replaced_vers, + postponed_cfgs); + + pk = package_key (bp->db, bp->name ()); + + // Note that here we side-step the bogus logic (by not setting + // the skipped flag) because in this case (replace=true) our + // choices are either (potentially) bogus or pathological + // (where we have evaluated too far). In other words, the + // postponed entry may cause the depends entry that triggered + // it to disappear (and thus, strictly speaking, to become + // bogus) but if we cancel it, we will be back to square one. + } + } + } + + // Make sure that this existing dependent doesn't belong to any + // (being) negotiated configuration cluster with a greater + // dependency index. That would mean that this dependent has already + // been re-evaluated to this index and so cannot participate in the + // configuration negotiation of this earlier dependency. + // + for (const postponed_configuration& cfg: postponed_cfgs) + { + if (const pair* p = + cfg.existing_dependent_position (pk)) + { + size_t ei (p->first); + + if (di < ei && cfg.negotiated) + { + // Feels like there cannot be an earlier position. + // + postponed_position pp (ed.dependency_position, + false /* replace */); + + auto p (postponed_poss.emplace (move (pk), pp)); + if (!p.second) + { + assert (p.first->second > pp); + p.first->second = pp; + } + + l5 ([&]{trace << "cannot cfg-postpone dependency " + << bp->available_name_version_db () + << " of existing dependent " << *ed.selected + << ed.db << " (index " << di + << ") due to earlier dependency index " << ei + << " in " << cfg << ", throwing " + << "postpone_position";}); + + // Don't print the "while satisfying..." chain. + // + dep_chain.clear (); + + throw postpone_position (); + } + + if (di == ei) + { + // For the negotiated cluster all the dependency packages + // should have been added. For non-negotiated cluster we + // cannot add the missing dependencies at the moment and will + // do it as a part of the dependent re-evaluation. + // + assert (!cfg.negotiated); + } + } + } + + l5 ([&]{trace << "cfg-postpone dependency " + << bp->available_name_version_db () + << " of existing dependent " << *ed.selected + << ed.db;}); + + postponed_cfgs.add (move (dpk), ed.dependency_position, move (pk)); + } + + return; + } + } + + pkg.recursive_collection = true; + + if (pkg.system) + { + l5 ([&]{trace << "skip system " << pkg.available_name_version_db ();}); + return; + } + + const shared_ptr& ap (pkg.available); + assert (ap != nullptr); + + const shared_ptr& sp (pkg.selected); + + // True if this is an up/down-grade. + // + bool ud (sp != nullptr && sp->version != pkg.available_version ()); + + // If this is a repointed dependent, then it points to its prerequisite + // replacements flag map (see repointed_dependents for details). + // + const map* rpt_prereq_flags (nullptr); + + // Bail out if this is a configured non-system package and no recursive + // collection is required. + // + bool src_conf (sp != nullptr && + sp->state == package_state::configured && + sp->substate != package_substate::system); + + // The being re-evaluated dependent must be configured as a source package + // and should not be collected recursively (due to upgrade, etc). + // + assert (!reeval || (src_conf && !pkg.recollect_recursively (rpt_depts))); + + if (src_conf) + { + repointed_dependents::const_iterator i (rpt_depts.find (pk)); + + if (i != rpt_depts.end ()) + rpt_prereq_flags = &i->second; + + if (!reeval && !pkg.recollect_recursively (rpt_depts)) + { + l5 ([&]{trace << "skip configured " + << pkg.available_name_version_db ();}); + return; + } + } + + // Iterate over dependencies, trying to unambiguously select a + // satisfactory dependency alternative for each of them. Fail or postpone + // the collection if unable to do so. + // + const dependencies& deps (ap->dependencies); + + // The skeleton can be pre-initialized before the recursive collection + // starts (as a part of dependency configuration negotiation, etc). The + // dependencies and alternatives members must both be either present or + // not. + // + assert ((!pkg.dependencies || pkg.skeleton) && + pkg.dependencies.has_value () == pkg.alternatives.has_value ()); + + // Note that the selected alternatives list can be filled partially (see + // build_package::dependencies for details). In this case we continue + // collecting where we stopped previously. + // + if (!pkg.dependencies) + { + l5 ([&]{trace << (reeval ? "reeval " : "begin ") + << pkg.available_name_version_db ();}); + + pkg.dependencies = dependencies (); + pkg.alternatives = vector (); + + if (size_t n = deps.size ()) + { + pkg.dependencies->reserve (n); + pkg.alternatives->reserve (n); + } + + if (!pkg.skeleton) + pkg.init_skeleton (options); + } + else + l5 ([&]{trace << "resume " << pkg.available_name_version_db ();}); + + dependencies& sdeps (*pkg.dependencies); + vector& salts (*pkg.alternatives); + + assert (sdeps.size () == salts.size ()); // Must be parallel. + + // Check if there is nothing to collect anymore. + // + if (sdeps.size () == deps.size ()) + { + l5 ([&]{trace << "end " << pkg.available_name_version_db ();}); + return; + } + + // Show how we got here if things go wrong. + // + // To suppress printing this information clear the dependency chain before + // throwing an exception. + // + auto g ( + make_exception_guard ( + [&dep_chain] () + { + // Note that we also need to clear the dependency chain, to prevent + // the caller's exception guard from printing it. + // + while (!dep_chain.empty ()) + { + info << "while satisfying " + << dep_chain.back ().get ().available_name_version_db (); + + dep_chain.pop_back (); + } + })); + + dep_chain.push_back (pkg); + + assert (sdeps.size () < deps.size ()); + + package_skeleton& skel (*pkg.skeleton); + + auto fail_reeval = [&pkg] () + { + fail << "unable to re-create dependency information of already " + << "configured package " << pkg.available_name_version_db () << + info << "likely cause is change in external environment" << + info << "consider resetting the build configuration"; + }; + + bool postponed (false); + bool reevaluated (false); + + for (size_t di (sdeps.size ()); di != deps.size (); ++di) + { + // Fail if we missed the re-evaluation target position for any reason. + // + if (reeval && di == reeval_pos.first) // Note: reeval_pos is 1-based. + fail_reeval (); + + const dependency_alternatives_ex& das (deps[di]); + + // Add an empty alternatives list into the selected dependency list if + // this is a toolchain build-time dependency. + // + dependency_alternatives_ex sdas (das.buildtime, das.comment); + + if (toolchain_buildtime_dependency (options, das, &nm)) + { + sdeps.push_back (move (sdas)); + salts.push_back (0); // Keep parallel to sdeps. + continue; + } + + // Evaluate alternative conditions and filter enabled alternatives. Add + // an empty alternatives list into the selected dependency list if there + // are none. + // + build_package::dependency_alternatives_refs edas; + + if (pkg.postponed_dependency_alternatives) + { + edas = move (*pkg.postponed_dependency_alternatives); + pkg.postponed_dependency_alternatives = nullopt; + } + else + { + for (size_t i (0); i != das.size (); ++i) + { + const dependency_alternative& da (das[i]); + + if (!da.enable || + skel.evaluate_enable (*da.enable, make_pair (di, i))) + edas.push_back (make_pair (ref (da), i)); + } + } + + if (edas.empty ()) + { + sdeps.push_back (move (sdas)); + salts.push_back (0); // Keep parallel to sdeps. + continue; + } + + // Try to pre-collect build information (pre-builds) for the + // dependencies of an alternative. Optionally, issue diagnostics into + // the specified diag record. In the dry-run mode don't change the + // packages collection state (postponed_repo set, etc). + // + // Note that rather than considering an alternative as unsatisfactory + // (returning no pre-builds) the function can fail in some cases + // (multiple possible configurations for a build-time dependency, orphan + // or broken selected package, etc). The assumption here is that the + // user would prefer to fix a dependency-related issue first instead of + // proceeding with the build which can potentially end up with some less + // preferable dependency alternative. + // + struct prebuild + { + bpkg::dependency dependency; + reference_wrapper db; + shared_ptr selected; + shared_ptr available; + lazy_shared_ptr repository_fragment; + bool system; + bool specified_dependency; + bool force; + + // True if the dependency package is either selected in the + // configuration or is already being built. + // + bool reused; + }; + using prebuilds = small_vector; + + class precollect_result + { + public: + // Nullopt if some dependencies cannot be resolved. + // + optional builds; + + // If some dependency of the alternative cannot be resolved because + // there is no version available which can satisfy all the being built + // dependents, then this member contains all the dependency builds + // (which otherwise would be contained in the builds member). + // + optional unsatisfactory; + + // True if dependencies can all be resolved (builds is present) and + // are all reused (see above). + // + bool reused = false; + + // True if some of the dependencies cannot be resolved (builds is + // nullopt) and the dependent package prerequisites collection needs + // to be postponed due to inability to find a version satisfying the + // pre-entered constraint from repositories available to the dependent + // package. + // + bool repo_postpone = false; + + // Create precollect result containing dependency builds. + // + precollect_result (prebuilds&& bs, bool r) + : builds (move (bs)), reused (r) {} + + // Create precollect result containing unsatisfactory dependency + // builds. + // + precollect_result (bool r, prebuilds&& bs) + : unsatisfactory (move (bs)), reused (r) {} + + // Create precollect result without builds (some dependency can't be + // resolved, etc). + // + explicit + precollect_result (bool p): repo_postpone (p) {} + }; + + auto precollect = [&options, + &pkg, + &nm, + &pdb, + ud, + &fdb, + rpt_prereq_flags, + &apc, + postponed_repo, + &dep_chain, + &trace, + this] + (const dependency_alternative& da, + bool buildtime, + const package_prerequisites* prereqs, + diag_record* dr = nullptr, + bool dry_run = false) -> precollect_result + { + prebuilds r; + bool reused (true); + + const lazy_shared_ptr& af ( + pkg.repository_fragment); + + for (const dependency& dp: da) + { + const package_name& dn (dp.name); + + if (buildtime && pdb.type == build2_config_type) + { + assert (dr == nullptr); // Should fail on the "silent" run. + + // Note that the dependent is not necessarily a build system + // module. + // + fail << "build-time dependency " << dn << " in build system " + << "module configuration" << + info << "build system modules cannot have build-time " + << "dependencies"; + } + + bool system (false); + bool specified (false); + + // If the user specified the desired dependency version + // constraint, then we will use it to overwrite the constraint + // imposed by the dependent package, checking that it is still + // satisfied. + // + // Note that we can't just rely on the execution plan refinement + // that will pick up the proper dependency version at the end of + // the day. We may just not get to the plan execution simulation, + // failing due to inability for dependency versions collected by + // two dependents to satisfy each other constraints (for an + // example see the + // pkg-build/dependency/apply-constraints/resolve-conflict/ + // tests). + + // Points to the desired dependency version constraint, if + // specified, and is NULL otherwise. Can be used as boolean flag. + // + const version_constraint* dep_constr (nullptr); + + database* ddb (fdb (pdb, dn, buildtime)); + + auto i (ddb != nullptr + ? map_.find (*ddb, dn) + : map_.find_dependency (pdb, dn, buildtime)); + + if (i != map_.end ()) + { + const build_package& bp (i->second.package); + + specified = !bp.action; // Is pre-entered. + + if (specified && + // + // The version constraint is specified, + // + bp.hold_version && *bp.hold_version) + { + assert (bp.constraints.size () == 1); + + const build_package::constraint_type& c (bp.constraints[0]); + + dep_constr = &c.value; + system = bp.system; + + // If the user-specified dependency constraint is the wildcard + // version, then it satisfies any dependency constraint. + // + if (!wildcard (*dep_constr) && + !satisfies (*dep_constr, dp.constraint)) + { + if (dr != nullptr) + *dr << error << "unable to satisfy constraints on package " + << dn << + info << nm << pdb << " depends on (" << dn + << " " << *dp.constraint << ")" << + info << c.dependent << c.db << " depends on (" << dn + << " " << c.value << ")" << + info << "specify " << dn << " version to satisfy " << nm + << " constraint"; + + return precollect_result (false /* postpone */); + } + } + } + + const dependency& d (!dep_constr + ? dp + : dependency {dn, *dep_constr}); + + // First see if this package is already selected. If we already + // have it in the configuration and it satisfies our dependency + // version constraint, then we don't want to be forcing its + // upgrade (or, worse, downgrade). + // + // If the prerequisite configuration is explicitly specified by + // the user, then search for the prerequisite in this specific + // configuration. Otherwise, search recursively in the explicitly + // linked configurations of the dependent configuration. + // + // Note that for the repointed dependent we will always find the + // prerequisite replacement rather than the prerequisite being + // replaced. + // + pair, database*> spd ( + ddb != nullptr + ? make_pair (ddb->find (dn), ddb) + : find_dependency (pdb, dn, buildtime)); + + if (ddb == nullptr) + ddb = &pdb; + + shared_ptr& dsp (spd.first); + + if (prereqs != nullptr && + (dsp == nullptr || + find_if (prereqs->begin (), prereqs->end (), + [&dsp] (const auto& v) + { + return v.first.object_id () == dsp->name; + }) == prereqs->end ())) + return precollect_result (false /* postpone */); + + pair, + lazy_shared_ptr> rp; + + shared_ptr& dap (rp.first); + + bool force (false); + + if (dsp != nullptr) + { + // Switch to the selected package configuration. + // + ddb = spd.second; + + // If we are collecting prerequisites of the repointed + // dependent, then only proceed further if this is either a + // replacement or unamended prerequisite and we are + // up/down-grading (only for the latter). + // + if (rpt_prereq_flags != nullptr) + { + auto i (rpt_prereq_flags->find (package_key {*ddb, dn})); + + bool unamended (i == rpt_prereq_flags->end ()); + bool replacement (!unamended && i->second); + + // We can never end up with the prerequisite being replaced, + // since the fdb() function should always return the + // replacement instead (see above). + // + assert (unamended || replacement); + + if (!(replacement || (unamended && ud))) + continue; + } + + if (dsp->state == package_state::broken) + { + assert (dr == nullptr); // Should fail on the "silent" run. + + fail << "unable to build broken package " << dn << *ddb << + info << "use 'pkg-purge --force' to remove"; + } + + // If the constraint is imposed by the user we also need to make + // sure that the system flags are the same. + // + if (satisfies (dsp->version, d.constraint) && + (!dep_constr || dsp->system () == system)) + { + system = dsp->system (); + + version_constraint vc (dsp->version); + + // First try to find an available package for this exact + // version, falling back to ignoring version revision and + // iteration. In particular, this handles the case where a + // package moves from one repository to another (e.g., from + // testing to stable). For a system package we will try to + // find the available package that matches the selected + // package version (preferable for the configuration + // negotiation machinery) and, if fail, fallback to picking + // the latest one (its exact version doesn't really matter in + // this case). + // + // It seems reasonable to search for the package in the + // repositories explicitly added by the user if the selected + // package was explicitly specified on command line, and in + // the repository (and its complements/prerequisites) of the + // dependent being currently built otherwise. + // + if (dsp->hold_package) + { + linked_databases dbs (dependent_repo_configs (*ddb)); + + rp = find_available_one (dbs, + dn, + vc, + true /* prereq */, + true /* revision */); + + if (dap == nullptr) + rp = find_available_one (dbs, dn, vc); + + if (dap == nullptr && system) + rp = find_available_one (dbs, dn, nullopt); + } + else if (af != nullptr) + { + rp = find_available_one (dn, + vc, + af, + true /* prereq */, + true /* revision */); + + if (dap == nullptr) + rp = find_available_one (dn, vc, af); + + if (dap == nullptr && system) + rp = find_available_one (dn, nullopt, af); + } + + // A stub satisfies any version constraint so we weed them out + // (returning stub as an available package feels wrong). + // + if (dap == nullptr || dap->stub ()) + rp = make_available_fragment (options, *ddb, dsp); + } + else + // Remember that we may be forcing up/downgrade; we will deal + // with it below. + // + force = true; + } + + // If this is a build-time dependency and we build it for the + // first time, then we need to find a suitable configuration (of + // the host or build2 type) to build it in. + // + // If the current configuration (ddb) is of the suitable type, + // then we use that. Otherwise, we go through its immediate + // explicit links. If only one of them has the suitable type, then + // we use that. If there are multiple of them, then we fail + // advising the user to pick one explicitly. If there are none, + // then we create the private configuration and use that. If the + // current configuration is private, then search/create in the + // parent configuration instead. + // + // Note that if the user has explicitly specified the + // configuration for this dependency on the command line (using + // --config-*), then this configuration is used as the starting + // point for this search. + // + if (buildtime && + dsp == nullptr && + ddb->type != buildtime_dependency_type (dn)) + { + database* db (nullptr); + database& sdb (ddb->private_ () ? ddb->parent_config () : *ddb); + + const string& type (buildtime_dependency_type (dn)); + + // Skip the self-link. + // + const linked_configs& lcs (sdb.explicit_links ()); + for (auto i (lcs.begin_linked ()); i != lcs.end (); ++i) + { + database& ldb (i->db); + + if (ldb.type == type) + { + if (db == nullptr) + db = &ldb; + else + { + assert (dr == nullptr); // Should fail on the "silent" run. + + fail << "multiple possible " << type << " configurations " + << "for build-time dependency (" << dp << ")" << + info << db->config_orig << + info << ldb.config_orig << + info << "use --config-* to select the configuration"; + } + } + } + + // If no suitable configuration is found, then create and link + // it, unless the --no-private-config options is specified. In + // the latter case, print the dependency chain to stdout and + // exit with the specified code. + // + if (db == nullptr) + { + // The private config should be created on the "silent" run + // and so there always should be a suitable configuration on + // the diagnostics run. + // + assert (dr == nullptr); + + if (options.no_private_config_specified ()) + try + { + // Note that we don't have the dependency package version + // yet. We could probably rearrange the code and obtain the + // available dependency package by now, given that it comes + // from the main database and may not be specified as system + // (we would have the configuration otherwise). However, + // let's not complicate the code further and instead print + // the package name and the constraint, if present. + // + // Also, in the future, we may still need the configuration + // to obtain the available dependency package for some + // reason (may want to fetch repositories locally, etc). + // + cout << d << '\n'; + + // Note that we also need to clean the dependency chain, to + // prevent the exception guard from printing it to stderr. + // + for (build_package_refs dc (move (dep_chain)); + !dc.empty (); ) + { + const build_package& p (dc.back ()); + + cout << p.available_name_version () << ' ' + << p.db.get ().config << '\n'; + + dc.pop_back (); + } + + throw failed (options.no_private_config ()); + } + catch (const io_error&) + { + fail << "unable to write to stdout"; + } + + const strings mods {"cc"}; + + const strings vars { + "config.config.load=~" + type, + "config.config.persist+='config.*'@unused=drop"}; + + dir_path cd (bpkg_dir / dir_path (type)); + + // Wipe a potentially existing un-linked private configuration + // left from a previous faulty run. Note that trying to reuse + // it would be a bad idea since it can be half-prepared, with + // an outdated database schema version, etc. + // + cfg_create (options, + sdb.config_orig / cd, + optional (type) /* name */, + type /* type */, + mods, + vars, + false /* existing */, + true /* wipe */); + + // Note that we will copy the name from the configuration + // unless it clashes with one of the existing links. + // + shared_ptr lc ( + cfg_link (sdb, + sdb.config / cd, + true /* relative */, + nullopt /* name */, + true /* sys_rep */)); + + // Save the newly-created private configuration, together with + // the containing configuration database, for their subsequent + // re-link. + // + apc (sdb, move (cd)); + + db = &sdb.find_attached (*lc->id); + } + + ddb = db; // Switch to the dependency configuration. + } + + // Note that building a dependent which is not a build2 module in + // the same configuration with the build2 module it depends upon + // is an error. + // + if (buildtime && + !build2_module (nm) && + build2_module (dn) && + pdb == *ddb) + { + assert (dr == nullptr); // Should fail on the "silent" run. + + // Note that the dependent package information is printed by the + // above exception guard. + // + fail << "unable to build build system module " << dn + << " in its dependent package configuration " + << pdb.config_orig << + info << "use --config-* to select suitable configuration"; + } + + // If we didn't get the available package corresponding to the + // selected package, look for any that satisfies the constraint. + // + if (dap == nullptr) + { + // And if we have no repository fragment to look in, then that + // means the package is an orphan (we delay this check until we + // actually need the repository fragment to allow orphans + // without prerequisites). + // + if (af == nullptr) + { + assert (dr == nullptr); // Should fail on the "silent" run. + + fail << "package " << pkg.available_name_version_db () + << " is orphaned" << + info << "explicitly upgrade it to a new version"; + } + + // We look for prerequisites only in the repositories of this + // package (and not in all the repositories of this + // configuration). At first this might look strange, but it also + // kind of makes sense: we only use repositories "approved" for + // this package version. Consider this scenario as an example: + // hello/1.0.0 and libhello/1.0.0 in stable and libhello/2.0.0 + // in testing. As a prerequisite of hello, which version should + // libhello resolve to? While one can probably argue either way, + // resolving it to 1.0.0 is the conservative choice and the user + // can always override it by explicitly building libhello. + // + // Note though, that if this is a test package, then its special + // test dependencies (main packages that refer to it) should be + // searched upstream through the complement repositories + // recursively, since the test packages may only belong to the + // main package's repository and its complements. + // + // @@ Currently we don't implement the reverse direction search + // for the test dependencies, effectively only supporting the + // common case where the main and test packages belong to the + // same repository. Will need to fix this eventually. + // + // Note that this logic (naturally) does not apply if the + // package is already selected by the user (see above). + // + // Also note that for the user-specified dependency version + // constraint we rely on the satisfying package version be + // present in repositories of the first dependent met. As a + // result, we may fail too early if such package version doesn't + // belong to its repositories, but belongs to the ones of some + // dependent that we haven't met yet. Can we just search all + // repositories for an available package of the appropriate + // version and just take it, if present? We could, but then + // which repository should we pick? The wrong choice can + // introduce some unwanted repositories and package versions + // into play. So instead, we will postpone collecting the + // problematic dependent, expecting that some other one will + // find the appropriate version in its repositories. + // + // For a system package we will try to find the available + // package that matches the constraint (preferable for the + // configuration negotiation machinery) and, if fail, fallback + // to picking the latest one just to make sure the package is + // recognized. An unrecognized package means the broken/stale + // repository (see below). + // + rp = find_available_one (dn, d.constraint, af); + + if (dap == nullptr && system && d.constraint) + rp = find_available_one (dn, nullopt, af); + + if (dap == nullptr) + { + if (dep_constr && !system && postponed_repo != nullptr) + { + // We shouldn't be called in the diag mode for the postponed + // package builds. + // + assert (dr == nullptr); + + if (!dry_run) + { + l5 ([&]{trace << "rep-postpone dependent " + << pkg.available_name_version_db () + << " due to dependency " << dp + << " and user-specified constraint " + << *dep_constr;}); + + postponed_repo->insert (&pkg); + } + + return precollect_result (true /* postpone */); + } + + // Fail if we are unable to find an available dependency + // package which satisfies the dependent's constraint. + // + // It feels that just considering this alternative as + // unsatisfactory and silently trying another alternative + // would be wrong, since the user may rather want to + // fix/re-fetch the repository and retry. + // + diag_record dr (fail); + + // Issue diagnostics differently based on the presence of + // available packages for the unsatisfied dependency. + // + // Note that there can't be any stubs, since they satisfy + // any constraint and we won't be here if there were any. + // + vector> aps ( + find_available (dn, nullopt /* version_constraint */, af)); + + if (!aps.empty ()) + { + dr << "unable to satisfy dependency constraint (" << dn; + + // We need to be careful not to print the wildcard-based + // constraint. + // + if (d.constraint && + (!dep_constr || !wildcard (*dep_constr))) + dr << ' ' << *d.constraint; + + dr << ") of package " << nm << pdb << + info << "available " << dn << " versions:"; + + for (const shared_ptr& ap: aps) + dr << ' ' << ap->version; + } + else + { + dr << "no package available for dependency " << dn + << " of package " << nm << pdb; + } + + // Avoid printing this if the dependent package is external + // since it's more often confusing than helpful (they are + // normally not fetched manually). + // + if (!af->location.empty () && + !af->location.directory_based () && + (!dep_constr || system)) + dr << info << "repository " << af->location << " appears " + << "to be broken" << + info << "or the repository state could be stale" << + info << "run 'bpkg rep-fetch' to update"; + } + + // If all that's available is a stub then we need to make sure + // the package is present in the system repository and it's + // version satisfies the constraint. If a source package is + // available but there is a system package specified on the + // command line and it's version satisfies the constraint then + // the system package should be preferred. To recognize such a + // case we just need to check if the authoritative system + // version is set and it satisfies the constraint. If the + // corresponding system package is non-optional it will be + // preferred anyway. + // + if (dap->stub ()) + { + // Note that the constraint can safely be printed as it can't + // be a wildcard (produced from the user-specified dependency + // version constraint). If it were, then the system version + // wouldn't be NULL and would satisfy itself. + // + if (dap->system_version (*ddb) == nullptr) + { + if (dr != nullptr) + *dr << error << "dependency " << d << " of package " + << nm << " is not available in source" << + info << "specify ?sys:" << dn << " if it is available " + << "from the system"; + + return precollect_result (false /* postpone */); + } + + if (!satisfies (*dap->system_version (*ddb), d.constraint)) + { + if (dr != nullptr) + *dr << error << "dependency " << d << " of package " + << nm << " is not available in source" << + info << package_string (dn, + *dap->system_version (*ddb), + true /* system */) + << " does not satisfy the constrains"; + + return precollect_result (false /* postpone */); + } + + system = true; + } + else + { + auto p (dap->system_version_authoritative (*ddb)); + + if (p.first != nullptr && + p.second && // Authoritative. + satisfies (*p.first, d.constraint)) + system = true; + } + } + + bool ru (i != map_.end () || dsp != nullptr); + + if (!ru) + reused = false; + + r.push_back (prebuild {d, + *ddb, + move (dsp), + move (dap), + move (rp.second), + system, + specified, + force, + ru}); + } + + // Now, as we have pre-collected the dependency builds, go through + // them and check that for those dependencies which are already + // being built we will be able to choose one of them (either + // existing or new) which satisfies all the dependents. If that's + // not the case, then issue the diagnostics, if requested, and + // return the unsatisfactory dependency builds. + // + // Note that collect_build() also performs this check but postponing + // it till then can end up in failing instead of selecting some + // other dependency alternative. + // + for (const prebuild& b: r) + { + const shared_ptr& dap (b.available); + + assert (dap != nullptr); // Otherwise we would fail earlier. + + const dependency& d (b.dependency); + + auto i (map_.find (b.db, d.name)); + + if (i != map_.end () && d.constraint) + { + const build_package& bp (i->second.package); + + if (bp.action && *bp.action == build_package::build) + { + const version& v1 (b.system + ? *dap->system_version (b.db) + : dap->version); + + const version& v2 (bp.available_version ()); + + if (v1 != v2) + { + using constraint_type = build_package::constraint_type; + + constraint_type c1 {pdb, nm.string (), *d.constraint}; + + if (!satisfies (v2, c1.value)) + { + for (const constraint_type& c2: bp.constraints) + { + if (!satisfies (v1, c2.value)) + { + if (dr != nullptr) + { + const package_name& n (d.name); + const string& d1 (c1.dependent); + const string& d2 (c2.dependent); + + *dr << error << "unable to satisfy constraints on " + << "package " << n << + info << d2 << c2.db << " depends on (" << n << ' ' + << c2.value << ")" << + info << d1 << c1.db << " depends on (" << n << ' ' + << c1.value << ")" << + info << "available " + << bp.available_name_version () << + info << "available " + << package_string (n, v1, b.system) << + info << "explicitly specify " << n << " version " + << "to manually satisfy both constraints"; + } + + return precollect_result (reused, move (r)); + } + } + } + } + } + } + } + + return precollect_result (move (r), reused); + }; + + // Try to collect the previously collected pre-builds. + // + // Return false if the dependent has configuration clauses and is + // postponed until dependencies configuration negotiation. + // + auto collect = [&options, + &pkg, + &pdb, + &nm, + &pk, + &fdb, + &rpt_depts, + &apc, + initial_collection, + &replaced_vers, + &dep_chain, + postponed_repo, + postponed_alts, + &postponed_deps, + &postponed_cfgs, + &postponed_poss, + &unacceptable_alts, + &di, + reeval, + &reeval_pos, + &reevaluated, + &fail_reeval, + &edas, + &das, + &precollect, + &trace, + this] + (const dependency_alternative& da, + size_t dai, + prebuilds&& bs, + const package_prerequisites* prereqs) + { + // Dependency alternative position. + // + pair dp (di + 1, dai + 1); + + if (reeval && + dp.first == reeval_pos.first && + dp.second != reeval_pos.second) + fail_reeval (); + + postponed_configuration::packages cfg_deps; + + for (prebuild& b: bs) + { + build_package bp { + build_package::build, + b.db, + b.selected, + b.available, + move (b.repository_fragment), + nullopt, // Dependencies. + nullopt, // Dependencies alternatives. + nullopt, // Package skeleton. + nullopt, // Postponed dependency alternatives. + false, // Recursive collection. + nullopt, // Hold package. + nullopt, // Hold version. + {}, // Constraints. + b.system, + false, // Keep output directory. + false, // Disfigure (from-scratch reconf). + false, // Configure-only. + nullopt, // Checkout root. + false, // Checkout purge. + strings (), // Configuration variables. + {pk}, // Required by (dependent). + true, // Required by dependents. + 0}; // State flags. + + const optional& constraint ( + b.dependency.constraint); + + // Add our constraint, if we have one. + // + // Note that we always add the constraint implied by the + // dependent. The user-implied constraint, if present, will be + // added when merging from the pre-entered entry. So we will have + // both constraints for completeness. + // + if (constraint) + bp.constraints.emplace_back (pdb, nm.string (), *constraint); + + // Now collect this prerequisite. If it was actually collected + // (i.e., it wasn't already there) and we are forcing a downgrade + // or upgrade, then refuse for a held version, warn for a held + // package, and print the info message otherwise, unless the + // verbosity level is less than two. + // + // Note though that while the prerequisite was collected it could + // have happen because it is an optional package and so not being + // pre-collected earlier. Meanwhile the package was specified + // explicitly and we shouldn't consider that as a + // dependency-driven up/down-grade enforcement. + // + // Here is an example of the situation we need to handle properly: + // + // repo: foo/2(->bar/2), bar/0+1 + // build sys:bar/1 + // build foo ?sys:bar/2 + // + // Pass the function which verifies we don't try to force + // up/downgrade of the held version and makes sure we don't print + // the dependency chain if replace_version will be thrown. + // + // Also note that we rely on "small function object" optimization + // here. + // + struct + { + const build_package& dependent; + const prebuild& prerequisite; + } dpn {pkg, b}; + + const function verify ( + [&dpn, &dep_chain] (const build_package& p, bool scratch) + { + const prebuild& prq (dpn.prerequisite); + const build_package& dep (dpn.dependent); + + if (prq.force && !prq.specified_dependency) + { + // Fail if the version is held. Otherwise, warn if the + // package is held. + // + bool f (prq.selected->hold_version); + bool w (!f && prq.selected->hold_package); + + // Note that there is no sense to warn or inform the user if + // we are about to start re-collection from scratch. + // + // @@ It seems that we may still warn/inform multiple times + // about the same package if we start from scratch. The + // intermediate diagnostics can probably be irrelevant to + // the final result. + // + // Perhaps what we should do is queue the diagnostics and + // then, if the run is not scratched, issues it. And if + // it is scratched, then drop it. + // + if (f || ((w || verb >= 2) && !scratch)) + { + const version& av (p.available_version ()); + + bool u (av > prq.selected->version); + bool c (prq.dependency.constraint); + + diag_record dr; + + (f ? dr << fail : + w ? dr << warn : + dr << info) + << "package " << dep.name () << dep.db + << " dependency on " << (c ? "(" : "") << prq.dependency + << (c ? ")" : "") << " is forcing " + << (u ? "up" : "down") << "grade of " << *prq.selected + << prq.db << " to "; + + // Print both (old and new) package names in full if the + // system attribution changes. + // + if (prq.selected->system ()) + dr << p.available_name_version (); + else + dr << av; // Can't be a system version so is never wildcard. + + if (prq.selected->hold_version) + dr << info << "package version " << *prq.selected + << prq.db<< " is held"; + + if (f) + dr << info << "explicitly request version " + << (u ? "up" : "down") << "grade to continue"; + } + } + + // Don't print the "while satisfying..." chain if we are about + // to re-collect the packages. + // + if (scratch) + dep_chain.clear (); + }); + + // Note: non-recursive. + // + build_package* p ( + collect_build (options, + move (bp), + fdb, + rpt_depts, + apc, + initial_collection, + replaced_vers, + postponed_cfgs, + nullptr /* dep_chain */, + nullptr /* postponed_repo */, + nullptr /* postponed_alts */, + nullptr /* postponed_deps */, + nullptr /* postponed_poss */, + nullptr /* unacceptable_alts */, + verify)); + + package_key dpk (b.db, b.available->id.name); + + // Do not collect prerequisites recursively for dependent + // re-evaluation. Instead, if the re-evaluation position is + // reached, collect the dependency packages to add them to the + // existing dependent's cluster. + // + if (reeval) + { + if (dp == reeval_pos) + cfg_deps.push_back (move (dpk)); + + continue; + } + + // Do not recursively collect a dependency of a dependent with + // configuration clauses, which could be this or some other + // (indicated by the presence in postponed_deps) dependent. In the + // former case if the prerequisites were prematurely collected, + // throw postpone_dependency. + // + // Note that such a dependency will be recursively collected + // directly right after the configuration negotiation (rather than + // via the dependent). + // + bool collect_prereqs (p != nullptr); + + { + build_package* bp (entered_build (dpk)); + assert (bp != nullptr); + + if (da.prefer || da.require) + { + // Indicate that the dependent with configuration clauses is + // present. + // + { + auto i (postponed_deps.find (dpk)); + + // Do not override postponements recorded during postponed + // collection phase with those recorded during initial + // phase. + // + if (i == postponed_deps.end ()) + { + postponed_deps.emplace (dpk, + postponed_dependency { + false /* without_config */, + true /* with_config */, + initial_collection}); + } + else + i->second.with_config = true; + } + + // Prematurely collected before we saw any config clauses. + // + if (bp->recursive_collection && + postponed_cfgs.find_dependency (dpk) == nullptr) + { + l5 ([&]{trace << "cannot cfg-postpone dependency " + << bp->available_name_version_db () + << " of dependent " + << pkg.available_name_version_db () + << " (collected prematurely), " + << "throwing postpone_dependency";}); + + // Don't print the "while satisfying..." chain. + // + dep_chain.clear (); + + throw postpone_dependency (move (dpk)); + } + + // Postpone until (re-)negotiation. + // + l5 ([&]{trace << "cfg-postpone dependency " + << bp->available_name_version_db () + << " of dependent " + << pkg.available_name_version_db ();}); + + cfg_deps.push_back (move (dpk)); + + collect_prereqs = false; + } + else + { + // Indicate that the dependent without configuration clauses + // is also present. + // + auto i (postponed_deps.find (dpk)); + if (i != postponed_deps.end ()) + { + l5 ([&]{trace << "dep-postpone dependency " + << bp->available_name_version_db () + << " of dependent " + << pkg.available_name_version_db ();}); + + i->second.wout_config = true; + + collect_prereqs = false; + } + else + { + l5 ([&]{trace << "no cfg-clause for dependency " + << bp->available_name_version_db () + << " of dependent " + << pkg.available_name_version_db ();}); + } + } + } + + if (collect_prereqs) + collect_build_prerequisites (options, + *p, + fdb, + rpt_depts, + apc, + initial_collection, + replaced_vers, + dep_chain, + postponed_repo, + postponed_alts, + 0 /* max_alt_index */, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + } + + // If this dependent has any dependencies with configurations + // clauses, then we need to deal with that. + // + // This is what we refer to as the "up-negotiation" where we + // negotiate the configuration of dependents that could not be + // postponed and handled all at once during "initial negotiation" in + // collect_build_postponed(). + // + if (!cfg_deps.empty ()) + { + // First, determine if there is any unprocessed reused dependency + // alternative that we can potentially use instead of the current + // one if it turns out that a configuration for some of its + // dependencies cannot be negotiated between all the dependents + // (see unacceptable_alternatives for details). + // + bool has_alt (false); + { + // Find the index of the current dependency alternative. + // + size_t i (0); + for (; i != edas.size (); ++i) + { + if (&edas[i].first.get () == &da) + break; + } + + // The current dependency alternative must be present in the + // list. + // + assert (i != edas.size ()); + + // Return true if the current alternative is unacceptable. + // + auto unacceptable = + [&pk, &pkg, di, &i, &edas, &unacceptable_alts] () + { + // Convert to 1-base. + // + pair pos (di + 1, edas[i].second + 1); + + return unacceptable_alts.find ( + unacceptable_alternative (pk, + pkg.available->version, + pos)) != + unacceptable_alts.end (); + }; + + // See if there is any unprocessed reused alternative to the + // right. + // + // Note that this is parallel to the alternative selection + // logic. + // + for (++i; i != edas.size (); ++i) + { + if (unacceptable ()) + continue; + + const dependency_alternative& a (edas[i].first); + + precollect_result r (precollect (a, + das.buildtime, + prereqs, + nullptr /* diag_record */, + true /* dru_run */)); + + if (r.builds && r.reused) + { + has_alt = true; + break; + } + } + + // If there are none and we are in the "recreate dependency + // decisions" mode, then repeat the search in the "make + // dependency decisions" mode. + // + if (!has_alt && prereqs != nullptr) + { + for (i = 0; i != edas.size (); ++i) + { + if (unacceptable ()) + continue; + + const dependency_alternative& a (edas[i].first); + + if (&a != &da) // Skip the current dependency alternative. + { + precollect_result r (precollect (a, + das.buildtime, + nullptr /* prereqs */, + nullptr /* diag_record */, + true /* dru_run */)); + + if (r.builds && r.reused) + { + has_alt = true; + break; + } + } + } + } + } + + // Re-evaluation is a special case (it happens during cluster + // negotiation; see collect_build_postponed()). + // + if (reeval) + { + reevaluated = true; + + // Note: the dependent may already exist in the cluster with a + // subset of dependencies. + // + postponed_configuration& cfg ( + postponed_cfgs.add (pk, + true /* existing */, + dp, + cfg_deps, + has_alt).first); + + // Can we merge clusters as a result? Seems so. + // + // - Simple case is if the cluster(s) being merged are not + // negotiated. Then perhaps we could handle this via the same + // logic that handles the addition of extra dependencies. + // + // - For the complex case, perhaps just making the resulting + // cluster shadow and rolling back, just like in the other + // case (non-existing dependent). + // + // Note: this is a special case of the below more general logic. + // + // Also note that we can distinguish the simple case by the fact + // that the resulting cluster is not negotiated. Note however, + // that in this case it is guaranteed that all the involved + // clusters will be merged into the cluster which the being + // re-evaluated dependent belongs to since this cluster (while + // not being negotiated) already has non-zero depth (see + // collect_build_postponed() for details). + // + assert (cfg.depth != 0); + + if (cfg.negotiated) + { + l5 ([&]{trace << "re-evaluating dependent " + << pkg.available_name_version_db () + << " involves negotiated configurations and " + << "results in " << cfg << ", throwing " + << "merge_configuration";}); + + // Don't print the "while satisfying..." chain. + // + dep_chain.clear (); + + throw merge_configuration {cfg.depth}; + } + + l5 ([&]{trace << "re-evaluating dependent " + << pkg.available_name_version_db () + << " results in " << cfg;}); + + return false; + } + + // As a first step add this dependent/dependencies to one of the + // new/existing postponed_configuration clusters, which could + // potentially cause some of them to be merged. Here are the + // possibilities and what we should do in each case. + // + // 1. Got added to a new cluster -- this dependent got postponed + // and we return false. + // + // 2. Got added to an existing non-yet-negotiated cluster (which + // could potentially involve merging a bunch of them) -- ditto. + // + // 3. Got added to an existing already-[being]-negotiated cluster + // (which could potentially involve merging a bunch of them, + // some negotiated, some being negotiated, and some not yet + // negotiated) -- see below logic. + // + // Note that if a dependent is postponed, it will be recursively + // recollected right after the configuration negotiation. + + // Note: don't move the argument from since may be needed for + // constructing exception. + // + pair> r ( + postponed_cfgs.add (pk, + false /* existing */, + dp, + cfg_deps, + has_alt)); + + postponed_configuration& cfg (r.first); + + if (cfg.depth == 0) + return false; // Cases (1) or (2). + else + { + // Case (3). + // + // There is just one complication: + // + // If all the merged clusters are already negotiated, then all + // is good: all the dependencies in cfg_deps have been collected + // recursively as part of the configuration negotiation (because + // everything in this cluster is already negotiated) and we can + // return true (no need to postpone any further steps). + // + // But if we merged clusters not yet negotiated, or, worse, + // being in the middle of negotiation, then we need to get this + // merged cluster into the fully negotiated state. The way we do + // it is by throwing merge_configuration (see below). + // + // When we are back here after throwing merge_configuration, + // then all the clusters have been pre-merged and our call to + // add() shouldn't have added any new cluster. In this case the + // cluster can either be already negotiated or being negotiated + // and we can proceed as in the "everything is negotiated case" + // above (we just need to get the the dependencies that we care + // about into the recursively collected state). + // + + // To recap, r.second values mean: + // + // absent -- shadow cluster-based merge is/being negotiated + // false -- some non or being negotiated + // true -- all have been negotiated + // + if (r.second && !*r.second) + { + // The partially negotiated case. + // + // Handling this in a straightforward way is not easy due to + // the being negotiated cases -- we have code up the stack + // that is in the middle of the negotiation logic. + // + // Another idea is to again throw to the outer try/catch frame + // (thus unwinding all the being negotiated code) and complete + // the work there. The problem with this approach is that + // without restoring the state we may end up with unrelated + // clusters that will have no corresponding try-catch frames + // (because we may unwind them in the process). + // + // So the approach we will use is the "shadow" idea for + // merging clusters. Specifically, we throw + // merge_configuration to the outer try/catch. At the catch + // site we make the newly merged cluster a shadow of the + // restored cluster and retry the same steps similar to + // retry_configuration. As we redo these steps, we consult the + // shadow cluster and if the dependent/dependency entry is + // there, then instead of adding it to another (new/existing) + // cluster that would later be merged into this non-shadow + // cluster, we add it directly to the non-shadow cluster + // (potentially merging other cluster which it feels like by + // definition should all be already fully negotiated). The end + // result is that once we reach this point again, there will + // be nothing to merge. + // + // The shadow check is part of postponed_configs::add(). + // + l5 ([&]{trace << "cfg-postponing dependent " + << pkg.available_name_version_db () + << " merges non-negotiated and/or being " + << "negotiated configurations in and results in " + << cfg << ", throwing merge_configuration";}); + + // Don't print the "while satisfying..." chain. + // + dep_chain.clear (); + + throw merge_configuration {cfg.depth}; + } + + // Up-negotiate the configuration and if it has changed, throw + // retry_configuration to the try/catch frame corresponding to + // the negotiation of the outermost merged cluster in order to + // retry the same steps (potentially refining the configuration + // as we go along) and likely (but not necessarily) ending up + // here again, at which point we up-negotiate again with the + // expectation that the configuration won't change (but if it + // does, then we throw again and do another refinement pass). + // + // In a sense, semantically, we should act like a one more + // iteration of the initial negotiation loop with the exception + // acting like a request to restart the refinement process from + // the beginning. + // + bool changed; + { + // Similar to initial negotiation, resolve package skeletons + // for this dependent and its dependencies. + // + assert (pkg.skeleton); + package_skeleton& dept (*pkg.skeleton); + + // If a dependency has already been recursively collected, + // then we can no longer call reload_defaults() or + // verify_sensible() on its skeleton. We could reset it, but + // then we wouldn't be able to continue using it if + // negotiate_configuration() below returns false. So it seems + // the most sensible approach is to make a temporary copy and + // reset that. + // + small_vector, 1> depcs; + forward_list depcs_storage; // Ref stability. + { + depcs.reserve (cfg_deps.size ()); + for (const package_key& pk: cfg_deps) + { + build_package* b (entered_build (pk)); + assert (b != nullptr); + + package_skeleton* depc; + if (b->recursive_collection) + { + assert (b->skeleton); + + depcs_storage.push_front (*b->skeleton); + depc = &depcs_storage.front (); + depc->reset (); + } + else + depc = &(b->skeleton + ? *b->skeleton + : b->init_skeleton (options)); + + depcs.push_back (*depc); + } + } + + optional c ( + negotiate_configuration ( + cfg.dependency_configurations, dept, dp, depcs, has_alt)); + + // If the dependency alternative configuration cannot be + // negotiated for this dependent, then add an entry to + // unacceptable_alts and throw unaccept_alternative to + // recollect from scratch. + // + if (!c) + { + unacceptable_alts.emplace (pk, pkg.available->version, dp); + + l5 ([&]{trace << "unable to cfg-negotiate dependency " + << "alternative " << dp.first << ',' + << dp.second << " for dependent " + << pkg.available_name_version_db () + << ", throwing unaccept_alternative";}); + + // Don't print the "while satisfying..." chain. + // + dep_chain.clear (); + + throw unaccept_alternative (); + } + else + changed = *c; + } + + // If the configuration hasn't changed, then we carry on. + // Otherwise, retry the negotiation from the beginning to refine + // the resulting configuration (see the catch block for + // retry_configuration). + // + if (changed) + { + l5 ([&]{trace << "cfg-postponing dependent " + << pkg.available_name_version_db () + << " involves (being) negotiated configurations " + << "and results in " << cfg + << ", throwing retry_configuration";}); + + // Don't print the "while satisfying..." chain. + // + dep_chain.clear (); + + throw retry_configuration {cfg.depth, move (pk)}; + } + + l5 ([&]{trace << "configuration for cfg-postponed " + << "dependencies of dependent " + << pkg.available_name_version_db () << " is " + << (r.second ? "" : "shadow-") << "negotiated";}); + + // Note that even in the fully negotiated case we may still add + // extra dependencies to this cluster which we still need to + // configure and recursively collect before indicating to the + // caller (returning true) that we are done with this depends + // value and the dependent is not postponed. + // + for (const package_key& p: cfg_deps) + { + build_package* b (entered_build (p)); + assert (b != nullptr); + + // Reconfigure the configured dependencies (see + // collect_build_postponed() for details). + // + if (b->selected != nullptr && + b->selected->state == package_state::configured) + b->flags |= build_package::adjust_reconfigure; + + if (!b->recursive_collection) + { + l5 ([&]{trace << "collecting cfg-postponed dependency " + << b->available_name_version_db () + << " of dependent " + << pkg.available_name_version_db ();}); + + // Similar to the inital negotiation case, verify and set + // the dependent configuration for this dependency. + // + { + assert (b->skeleton); // Should have been init'ed above. + + const package_configuration& pc ( + cfg.dependency_configurations[p]); + + pair pr (b->skeleton->available != nullptr + ? b->skeleton->verify_sensible (pc) + : make_pair (true, string ())); + + if (!pr.first) + { + diag_record dr (fail); + dr << "unable to negotiate sensible configuration for " + << "dependency " << p << '\n' + << " " << pr.second; + + dr << info << "negotiated configuration:\n"; + pc.print (dr, " "); + } + + b->skeleton->dependent_config (pc); + } + + collect_build_prerequisites (options, + *b, + fdb, + rpt_depts, + apc, + initial_collection, + replaced_vers, + dep_chain, + postponed_repo, + postponed_alts, + 0 /* max_alt_index */, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + } + else + l5 ([&]{trace << "dependency " + << b->available_name_version_db () + << " of dependent " + << pkg.available_name_version_db () + << " is already (being) recursively " + << "collected, skipping";}); + } + + return true; + } + } + + return true; + }; + + // Select a dependency alternative, copying it alone into the resulting + // dependencies list and evaluating its reflect clause, if present. + // + bool selected (false); + auto select = [&sdeps, &salts, &sdas, &skel, di, &selected] + (const dependency_alternative& da, size_t dai) + { + assert (sdas.empty ()); + + // Avoid copying enable/reflect not to evaluate them repeatedly. + // + sdas.emplace_back (nullopt /* enable */, + nullopt /* reflect */, + da.prefer, + da.accept, + da.require, + da /* dependencies */); + + sdeps.push_back (move (sdas)); + salts.push_back (dai); + + if (da.reflect) + skel.evaluate_reflect (*da.reflect, make_pair (di, dai)); + + selected = true; + }; + + // Postpone the prerequisite builds collection, optionally inserting the + // package to the postponements set (can potentially already be there) + // and saving the enabled alternatives. + // + auto postpone = [&pkg, &edas, &postponed] + (postponed_packages* postpones) + { + if (postpones != nullptr) + postpones->insert (&pkg); + + pkg.postponed_dependency_alternatives = move (edas); + postponed = true; + }; + + // Iterate over the enabled dependencies and try to select a + // satisfactory alternative. + // + // If the package is already configured as source and is not + // up/downgraded, then we will try to resolve its dependencies to the + // current prerequisites. To achieve this we will first try to select an + // alternative in the "recreate dependency decisions" mode, filtering + // out all the alternatives where dependencies do not all belong to the + // list of current prerequisites. If we end up with no alternative + // selected, then we retry in the "make dependency decisions" mode and + // select the alternative ignoring the current prerequisites. + // + // Note though, that if we are re-evaluating an existing dependent + // then we fail if we didn't succeed in the "recreate dependency + // decisions" mode. + // + const package_prerequisites* prereqs (src_conf && !ud + ? &sp->prerequisites + : nullptr); + + // During the dependent re-evaluation we always try to reproduce the + // existing setup. + // + assert (!reeval || prereqs != nullptr); + + for (bool unacceptable (false);;) + { + // The index and pre-collection result of the first satisfactory + // alternative. + // + optional> first_alt; + + // The number of satisfactory alternatives. + // + size_t alts_num (0); + + // If true, then only reused alternatives will be considered for the + // selection. + // + // The idea here is that we don't want to bloat the configuration by + // silently configuring a new dependency package as the alternative + // for an already used but not satisfactory for all the dependents + // dependency. Think of silently configuring Qt6 just because the + // configured version of Qt5 is not satisfactory for all the + // dependents. The user must have a choice if to either configure this + // new dependency by specifying it explicitly or, for example, to + // upgrade dependents so that the existing dependency is satisfactory + // for all of them. + // + // Note that if there are multiple alternatives with all their + // dependencies resolved/satisfied, then only reused alternatives are + // considered anyway. Thus, this flag only affects the single + // alternative case. + // + bool reused_only (false); + + for (size_t i (0); i != edas.size (); ++i) + { + // Skip the unacceptable alternatives. + // + { + // Convert to 1-base. + // + pair pos (di + 1, edas[i].second + 1); + + if (unacceptable_alts.find ( + unacceptable_alternative (pk, ap->version, pos)) != + unacceptable_alts.end ()) + { + unacceptable = true; + + l5 ([&]{trace << "dependency alternative " << pos.first << ',' + << pos.second << " for dependent " + << pkg.available_name_version_db () + << " is unacceptable, skipping";}); + + continue; + } + } + + const dependency_alternative& da (edas[i].first); + + precollect_result r (precollect (da, das.buildtime, prereqs)); + + // If we didn't come up with satisfactory dependency builds, then + // skip this alternative and try the next one, unless the collecting + // is postponed in which case just bail out. + // + // Should we skip alternatives for which we are unable to satisfy + // the constraint? On one hand, this could be a user error: there is + // no package available from dependent's repositories that satisfies + // the constraint. On the other hand, it could be that it's other + // dependent's constraints that we cannot satisfy together with + // others. And in this case we may want some other + // alternative. Consider, as an example, something like this: + // + // depends: libfoo >= 2.0.0 | {libfoo >= 1.0.0 libbar} + // + if (!r.builds) + { + if (r.repo_postpone) + { + if (reeval) + fail_reeval (); + + postpone (nullptr); // Already inserted into postponed_repo. + break; + } + + // If this alternative is reused but is not satisfactory, then + // switch to the reused-only mode. + // + if (r.reused && r.unsatisfactory) + reused_only = true; + + continue; + } + + ++alts_num; + + // Note that when we see the first satisfactory alternative, we + // don't know yet if it is a single alternative or the first of the + // (multiple) true alternatives (those are handled differently). + // Thus, we postpone its processing until the second satisfactory + // alternative is encountered or the end of the alternatives list is + // reached. + // + if (!first_alt) + { + first_alt = make_pair (i, move (r)); + continue; + } + + // Try to collect and then select a true alternative, returning true + // if the alternative is selected or the collection is postponed. + // Return false if the alternative is ignored (not postponed and not + // all of it dependencies are reused). + // + auto try_select = [postponed_alts, &max_alt_index, + &edas, &pkg, + prereqs, + reeval, + &trace, + &postpone, &collect, &select] + (size_t index, precollect_result&& r) + { + const auto& eda (edas[index]); + const dependency_alternative& da (eda.first); + size_t dai (eda.second); + + // Postpone the collection if the alternatives maximum index is + // reached. + // + if (postponed_alts != nullptr && index >= max_alt_index) + { + // For a dependent re-evaluation max_alt_index is expected to be + // max size_t. + // + assert (!reeval); + + l5 ([&]{trace << "alt-postpone dependent " + << pkg.available_name_version_db () + << " since max index is reached: " << index << + info << "dependency alternative: " << da;}); + + postpone (postponed_alts); + return true; + } + + // Select this alternative if all its dependencies are reused and + // do nothing about it otherwise. + // + if (r.reused) + { + // On the diagnostics run there shouldn't be any alternatives + // that we could potentially select. + // + assert (postponed_alts != nullptr); + + if (!collect (da, dai, move (*r.builds), prereqs)) + { + postpone (nullptr); // Already inserted into postponed_cfgs. + return true; + } + + select (da, dai); + + // Make sure no more true alternatives are selected during this + // function call unless we are re-evaluating a dependent. + // + if (!reeval) + max_alt_index = 0; + + return true; + } + else + return false; + }; + + // If we encountered the second satisfactory alternative, then this + // is the "multiple true alternatives" case. In this case we also + // need to process the first satisfactory alternative, which + // processing was delayed. + // + if (alts_num == 2) + { + assert (first_alt); + + if (try_select (first_alt->first, move (first_alt->second))) + break; + } + + if (try_select (i, move (r))) + break; + + // Not all of the alternative dependencies are reused, so go to + // the next alternative. + } + + // Bail out if the collection is postponed for any reason. + // + if (postponed) + break; + + // Select the single satisfactory alternative if it is reused or we + // are not in the reused-only mode. + // + if (!selected && alts_num == 1) + { + assert (first_alt); + + precollect_result& r (first_alt->second); + + assert (r.builds); + + if (r.reused || !reused_only) + { + // If there are any unacceptable alternatives, then the remaining + // one should be reused. + // + assert (!unacceptable || r.reused); + + const auto& eda (edas[first_alt->first]); + const dependency_alternative& da (eda.first); + size_t dai (eda.second); + + if (!collect (da, dai, move (*r.builds), prereqs)) + { + postpone (nullptr); // Already inserted into postponed_cfgs. + break; + } + + select (da, dai); + } + } + + // If an alternative is selected, then we are done. + // + if (selected) + break; + + // Fail or postpone the collection if no alternative is selected, + // unless we are re-evaluating a dependent or are in the "recreate + // dependency decisions" mode. In the latter case fail for + // re-evaluation and fall back to the "make dependency decisions" mode + // and retry otherwise. + // + if (prereqs != nullptr) + { + if (reeval) + fail_reeval (); + + prereqs = nullptr; + continue; + } + + // We shouldn't end up with the "no alternative to select" case if any + // alternatives are unacceptable. + // + assert (!unacceptable); + + // Issue diagnostics and fail if there are no satisfactory + // alternatives. + // + if (alts_num == 0) + { + diag_record dr; + for (const auto& da: edas) + precollect (da.first, das.buildtime, nullptr /* prereqs */, &dr); + + assert (!dr.empty ()); + + dr.flush (); + throw failed (); + } + + // Issue diagnostics and fail if there are multiple non-reused + // alternatives or there is a single non-reused alternative in the + // reused-only mode, unless the failure needs to be postponed. + // + assert (alts_num > (!reused_only ? 1 : 0)); + + if (postponed_alts != nullptr) + { + if (verb >= 5) + { + diag_record dr (trace); + dr << "alt-postpone dependent " + << pkg.available_name_version_db () + << " due to ambiguous alternatives"; + + for (const auto& da: edas) + dr << info << "alternative: " << da.first; + } + + postpone (postponed_alts); + break; + } + + diag_record dr (fail); + + dr << "unable to select dependency alternative for package " + << pkg.available_name_version_db () << + info << "explicitly specify dependency packages to manually " + << "select the alternative"; + + for (const auto& da: edas) + { + precollect_result r ( + precollect (da.first, das.buildtime, nullptr /* prereqs */)); + + if (r.builds) + { + assert (!r.reused); // We shouldn't be failing otherwise. + + dr << info << "alternative:"; + + // Only print the non-reused dependencies, which needs to be + // explicitly specified by the user. + // + for (const prebuild& b: *r.builds) + { + if (!b.reused) + dr << ' ' << b.dependency.name; + } + } + } + + // If there is only a single alternative (while we are in the + // reused-only mode), then also print the reused unsatisfactory + // alternatives and the reasons why they are not satisfactory. + // + if (alts_num == 1) + { + assert (reused_only); + + for (const auto& da: edas) + { + precollect_result r ( + precollect (da.first, das.buildtime, nullptr /* prereqs */)); + + if (r.reused && r.unsatisfactory) + { + // Print the alternative. + // + dr << info << "unsatisfactory alternative:"; + + for (const prebuild& b: *r.unsatisfactory) + dr << ' ' << b.dependency.name; + + // Print the reason. + // + precollect (da.first, + das.buildtime, + nullptr /* prereqs */, + &dr); + } + } + } + } + + if (postponed) + break; + } + + if (reeval) + { + if (!reevaluated) + fail_reeval (); + + assert (postponed); + } + + dep_chain.pop_back (); + + l5 ([&]{trace << (!postponed ? "end " : + reeval ? "re-evaluated " : + "postpone ") + << pkg.available_name_version_db ();}); + } + + void build_packages:: + collect_build_prerequisites (const pkg_build_options& o, + database& db, + const package_name& name, + const function& fdb, + const repointed_dependents& rpt_depts, + const function& apc, + bool initial_collection, + replaced_versions& replaced_vers, + postponed_packages& postponed_repo, + postponed_packages& postponed_alts, + size_t max_alt_index, + postponed_dependencies& postponed_deps, + postponed_configurations& postponed_cfgs, + postponed_positions& postponed_poss, + unacceptable_alternatives& unacceptable_alts) + { + auto mi (map_.find (db, name)); + assert (mi != map_.end ()); + + build_package_refs dep_chain; + + collect_build_prerequisites (o, + mi->second.package, + fdb, + rpt_depts, + apc, + initial_collection, + replaced_vers, + dep_chain, + &postponed_repo, + &postponed_alts, + max_alt_index, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + } + + void build_packages:: + collect_repointed_dependents (const pkg_build_options& o, + const repointed_dependents& rpt_depts, + replaced_versions& replaced_vers, + postponed_packages& postponed_repo, + postponed_packages& postponed_alts, + postponed_dependencies& postponed_deps, + postponed_configurations& postponed_cfgs, + postponed_positions& postponed_poss, + unacceptable_alternatives& unacceptable_alts, + const function& fdb, + const function& apc) + { + for (const auto& rd: rpt_depts) + { + database& db (rd.first.db); + const package_name& nm (rd.first.name); + + auto i (map_.find (db, nm)); + if (i != map_.end ()) + { + build_package& b (i->second.package); + + if (!b.action || *b.action != build_package::adjust) + { + if (!b.action || + (*b.action != build_package::drop && !b.reconfigure ())) + b.flags |= build_package::adjust_reconfigure; + + continue; + } + } + + shared_ptr sp (db.load (nm)); + + // The repointed dependent can be an orphan, so just create the + // available package from the selected package. + // + auto rp (make_available_fragment (o, db, sp)); + + // Add the prerequisite replacements as the required-by packages. + // + set required_by; + for (const auto& prq: rd.second) + { + if (prq.second) // Prerequisite replacement? + { + const package_key& pk (prq.first); + required_by.emplace (pk.db, pk.name); + } + } + + build_package p { + build_package::build, + db, + sp, + move (rp.first), + move (rp.second), + nullopt, // Dependencies. + nullopt, // Dependencies alternatives. + nullopt, // Package skeleton. + nullopt, // Postponed dependency alternatives. + false, // Recursive collection. + nullopt, // Hold package. + nullopt, // Hold version. + {}, // Constraints. + sp->system (), + false, // Keep output directory. + false, // Disfigure (from-scratch reconf). + false, // Configure-only. + nullopt, // Checkout root. + false, // Checkout purge. + strings (), // Configuration variables. + move (required_by), // Required by (dependencies). + false, // Required by dependents. + build_package::adjust_reconfigure | build_package::build_repoint}; + + build_package_refs dep_chain; + + // Note: recursive. + // + collect_build (o, + move (p), + fdb, + rpt_depts, + apc, + true /* initial_collection */, + replaced_vers, + postponed_cfgs, + &dep_chain, + &postponed_repo, + &postponed_alts, + &postponed_deps, + &postponed_poss, + &unacceptable_alts); + } + } + + void build_packages:: + collect_drop (const pkg_build_options& options, + database& db, + shared_ptr sp, + replaced_versions& replaced_vers) + { + tracer trace ("collect_drop"); + + package_key pk (db, sp->name); + + // If there is an entry for building specific version of the package (the + // available member is not NULL), then it wasn't created to prevent out + // drop (see replaced_versions for details). This rather mean that the + // replacement version is not being built anymore due to the plan + // refinement. Thus, just erase the entry in this case and continue. + // + auto vi (replaced_vers.find (pk)); + if (vi != replaced_vers.end () && !vi->second.replaced) + { + replaced_version& v (vi->second); + const shared_ptr& ap (v.available); + + if (ap != nullptr) + { + if (verb >= 5) + { + bool s (v.system); + const version& av (s ? *ap->system_version (db) : ap->version); + + l5 ([&]{trace << "erase version replacement for " + << package_string (ap->id.name, av, s) << db;}); + } + + replaced_vers.erase (vi); + vi = replaced_vers.end (); // Keep it valid for the below check. + } + else + v.replaced = true; + } + + build_package p { + build_package::drop, + db, + move (sp), + nullptr, + nullptr, + nullopt, // Dependencies. + nullopt, // Dependencies alternatives. + nullopt, // Package skeleton. + nullopt, // Postponed dependency alternatives. + false, // Recursive collection. + nullopt, // Hold package. + nullopt, // Hold version. + {}, // Constraints. + false, // System package. + false, // Keep output directory. + false, // Disfigure (from-scratch reconf). + false, // Configure-only. + nullopt, // Checkout root. + false, // Checkout purge. + strings (), // Configuration variables. + {}, // Required by. + false, // Required by dependents. + 0}; // State flags. + + auto i (map_.find (pk)); + + if (i != map_.end ()) + { + build_package& bp (i->second.package); + + if (bp.available != nullptr) + { + // Similar to the version replacement in collect_build(), see if + // in-place drop is possible (no dependencies, etc) and set scratch to + // false if that's the case. + // + bool scratch (true); + + // While checking if the package has any dependencies skip the + // toolchain build-time dependencies since they should be quite + // common. + // + if (!has_dependencies (options, bp.available->dependencies)) + scratch = false; + + l5 ([&]{trace << bp.available_name_version_db () + << " package version needs to be replaced " + << (!scratch ? "in-place " : "") << "with drop";}); + + if (scratch) + { + if (vi != replaced_vers.end ()) + vi->second = replaced_version (); + else + replaced_vers.emplace (move (pk), replaced_version ()); + + throw replace_version (); + } + } + + // Overwrite the existing (possibly pre-entered, adjustment, or repoint) + // entry. + // + l4 ([&]{trace << "overwrite " << pk;}); + + bp = move (p); + } + else + { + l4 ([&]{trace << "add " << pk;}); + + map_.emplace (move (pk), data_type {end (), move (p)}); + } + } + + void build_packages:: + collect_unhold (database& db, const shared_ptr& sp) + { + auto i (map_.find (db, sp->name)); + + // Currently, it must always be pre-entered. + // + assert (i != map_.end ()); + + build_package& bp (i->second.package); + + if (!bp.action) // Pre-entered. + { + build_package p { + build_package::adjust, + db, + sp, + nullptr, + nullptr, + nullopt, // Dependencies. + nullopt, // Dependencies alternatives. + nullopt, // Package skeleton. + nullopt, // Postponed dependency alternatives. + false, // Recursive collection. + nullopt, // Hold package. + nullopt, // Hold version. + {}, // Constraints. + sp->system (), + false, // Keep output directory. + false, // Disfigure (from-scratch reconf). + false, // Configure-only. + nullopt, // Checkout root. + false, // Checkout purge. + strings (), // Configuration variables. + {}, // Required by. + false, // Required by dependents. + build_package::adjust_unhold}; + + p.merge (move (bp)); + bp = move (p); + } + else + bp.flags |= build_package::adjust_unhold; + } + + void build_packages:: + collect_build_postponed (const pkg_build_options& o, + replaced_versions& replaced_vers, + postponed_packages& postponed_repo, + postponed_packages& postponed_alts, + postponed_dependencies& postponed_deps, + postponed_configurations& postponed_cfgs, + strings& postponed_cfgs_history, + postponed_positions& postponed_poss, + unacceptable_alternatives& unacceptable_alts, + const function& fdb, + const repointed_dependents& rpt_depts, + const function& apc, + postponed_configuration* pcfg) + { + // Snapshot of the package builds collection state. + // + // Note: should not include postponed_cfgs_history. + // + class snapshot + { + public: + snapshot (const build_packages& pkgs, + const postponed_packages& postponed_repo, + const postponed_packages& postponed_alts, + const postponed_dependencies& postponed_deps, + const postponed_configurations& postponed_cfgs) + : pkgs_ (pkgs), + postponed_deps_ (postponed_deps), + postponed_cfgs_ (postponed_cfgs) + { + auto save = [] (vector& d, const postponed_packages& s) + { + d.reserve (s.size ()); + + for (const build_package* p: s) + d.emplace_back (p->db, p->name ()); + }; + + save (postponed_repo_, postponed_repo); + save (postponed_alts_, postponed_alts); + } + + void + restore (build_packages& pkgs, + postponed_packages& postponed_repo, + postponed_packages& postponed_alts, + postponed_dependencies& postponed_deps, + postponed_configurations& postponed_cfgs) + { + pkgs = move (pkgs_); + postponed_cfgs = move (postponed_cfgs_); + postponed_deps = move (postponed_deps_); + + auto restore = [&pkgs] (postponed_packages& d, + const vector& s) + { + d.clear (); + + for (const package_key& p: s) + { + build_package* b (pkgs.entered_build (p)); + assert (b != nullptr); + d.insert (b); + } + }; + + restore (postponed_repo, postponed_repo_); + restore (postponed_alts, postponed_alts_); + } + + private: + // Note: try to use vectors instead of sets for storage to save + // memory. We could probably optimize this some more if necessary + // (there are still sets/maps inside). + // + build_packages pkgs_; + vector postponed_repo_; + vector postponed_alts_; + postponed_dependencies postponed_deps_; + postponed_configurations postponed_cfgs_; + }; + + // This exception is thrown if negotiation of the current cluster needs to + // be skipped until later. This is normally required if this cluster + // contains some existing dependent which needs to be re-evaluated to a + // dependency position greater than some other not yet negotiated cluster + // will re-evaluate this dependent to. Sometimes this another cluster yet + // needs to be created in which case the exception carries the information + // required for that (see the postponed_position's replace flag for + // details). + // + struct skip_configuration + { + optional dependent; + pair new_position; + + skip_configuration () = default; + + skip_configuration (existing_dependent&& d, pair n) + : dependent (move (d)), new_position (n) {} + }; + + size_t depth (pcfg != nullptr ? pcfg->depth : 0); + + string t ("collect_build_postponed (" + to_string (depth) + ")"); + tracer trace (t.c_str ()); + + string trace_suffix; + if (verb >= 5 && pcfg != nullptr) + { + trace_suffix += ' '; + trace_suffix += pcfg->string (); + } + + l5 ([&]{trace << "begin" << trace_suffix;}); + + if (pcfg != nullptr) + { + // This is what we refer to as the "initial negotiation" where we + // negotiate the configuration of dependents that could be postponed. + // Those that could not we "up-negotiate" in the collect() lambda of + // collect_build_prerequisites(). + // + using packages = postponed_configuration::packages; + + assert (!pcfg->negotiated); + + // Re-evaluate existing dependents with configuration clause for + // dependencies in this configuration cluster up to these + // dependencies. Omit dependents which are already being built or + // dropped. Note that these dependents, potentially with additional + // dependencies, will be added to this cluster with the `existing` flag + // as a part of the dependents' re-evaluation (see the collect lambda in + // collect_build_prerequisites() for details). + // + // After being re-evaluated the existing dependents are recursively + // collected in the same way as the new dependents. + // + { + // Map existing dependents to the dependencies they apply a + // configuration to. Also, collect the information which is required + // for a dependent re-evaluation and its subsequent recursive + // collection (selected package, etc). + // + // As mentioned earlier, we may end up adding additional dependencies + // to pcfg->dependencies which in turn may have additional existing + // dependents which we need to process. Feels like doing this + // iteratively is the best option. + // + // Note that we need to make sure we don't re-process the same + // existing dependents. + // + struct existing_dependent_ex: existing_dependent + { + packages dependencies; + bool reevaluated = false; + + existing_dependent_ex (existing_dependent&& ed) + : existing_dependent (move (ed)) {} + }; + map dependents; + + const packages& deps (pcfg->dependencies); + + // Note that the below collect_build_prerequisites() call can only add + // new dependencies to the end of the cluster's dependencies + // list. Thus on each iteration we will only add existing dependents + // of unprocessed/new dependencies. We will also skip the already + // re-evaluated existing dependents. + // + for (size_t i (0); i != deps.size (); ) + { + size_t n (dependents.size ()); + + for (; i != deps.size (); ++i) + { + // Note: this reference is only used while deps is unchanged. + // + const package_key& p (deps[i]); + + // If the dependent is being built, then check if it was + // re-evaluated to the position greater than the dependency + // position. Return true if that's the case, so this package is + // added to the resulting list and we can handle this situation. + // + // Note that we rely on "small function object" optimization + // here. + // + const function verify ( + [&postponed_cfgs, pcfg] + (const package_key& pk, pair pos) + { + for (const postponed_configuration& cfg: postponed_cfgs) + { + if (&cfg == pcfg || cfg.negotiated) + { + if (const pair* p = + cfg.existing_dependent_position (pk)) + { + if (p->first > pos.first) + return true; + } + } + } + + return false; + }); + + for (existing_dependent& ed: + query_existing_dependents (trace, + p.db, + p.name, + replaced_vers, + rpt_depts, + verify)) + { + package_key pk (ed.db, ed.selected->name); + + // If this dependent is present in postponed_deps, then it means + // someone depends on it with configuration and it's no longer + // considered an existing dependent (it will be reconfigured). + // However, this fact may not be reflected yet. And it can + // actually turn out bogus. + // + auto pi (postponed_deps.find (pk)); + if (pi != postponed_deps.end ()) + { + l5 ([&]{trace << "skip dep-postponed existing dependent " + << pk << " of dependency " << p;}); + + // Note that here we would re-evaluate the existing dependent + // without specifying any configuration for it. + // + pi->second.wout_config = true; + + continue; + } + + auto i (dependents.find (pk)); + size_t di (ed.dependency_position.first); + + // Skip re-evaluated dependent if the dependency index is + // greater than the one we have already re-evaluated to. If it + // is earlier, then add the entry to postponed_poss and throw + // postpone_position to recollect from scratch. Note that this + // entry in postponed_poss is with replacement. + // + if (i != dependents.end () && i->second.reevaluated) + { + size_t ci (i->second.dependency_position.first); + + if (di > ci) + continue; + + // The newly-introduced dependency must belong to the depends + // value other then the one we have re-evaluated to. + // + assert (di < ci); + + postponed_position pp (ed.dependency_position, + true /* replace */); + + auto p (postponed_poss.emplace (pk, pp)); + + if (!p.second) + { + assert (p.first->second > pp); + p.first->second = pp; + } + + l5 ([&]{trace << "cannot re-evaluate dependent " + << pk << " to dependency index " << di + << " since it is already re-evaluated to " + << "greater index " << ci << " in " << *pcfg + << ", throwing postpone_position";}); + + throw postpone_position (); + } + + // If the existing dependent is not in the map yet, then add + // it. Otherwise, if the dependency position is greater than + // that one in the existing map entry then skip it (this + // position will be up-negotiated, if it's still present). + // Otherwise, if the position is less then overwrite the + // existing entry. Otherwise (the position is equal), just add + // the dependency to the existing entry. + // + // Note that we want to re-evaluate the dependent up to the + // earliest dependency position and continue with the regular + // prerequisites collection (as we do for new dependents) + // afterwards. + // + if (i == dependents.end ()) + { + i = dependents.emplace ( + move (pk), existing_dependent_ex (move (ed))).first; + } + else + { + size_t ci (i->second.dependency_position.first); + + if (ci < di) + continue; + else if (ci > di) + i->second = existing_dependent_ex (move (ed)); + //else if (ci == di) + // ; + } + + i->second.dependencies.push_back (p); + } + } + + // Re-evaluate the newly added existing dependents, if any. + // + if (dependents.size () != n) + { + l5 ([&]{trace << "re-evaluate existing dependents for " << *pcfg;}); + + for (auto& d: dependents) + { + existing_dependent_ex& ed (d.second); + + // Skip re-evaluated. + // + if (ed.reevaluated) + continue; + + size_t di (ed.dependency_position.first); + const package_key& pk (d.first); + + // Check if there is an earlier dependency position for this + // dependent that will be participating in a configuration + // negotiation and skip this cluster if that's the case. There + // are two places to check: postponed_poss and other clusters. + // + auto pi (postponed_poss.find (pk)); + if (pi != postponed_poss.end () && pi->second.first < di) + { + l5 ([&]{trace << "pos-postpone existing dependent " + << pk << " re-evaluation to dependency " + << "index " << di << " due to recorded index " + << pi->second.first << ", skipping " << *pcfg;}); + + pi->second.skipped = true; + + // If requested, override the first encountered non-replace + // position to replace (see below for details). + // + if (!pi->second.replace && postponed_poss.replace) + { + pi->second.replace = true; + postponed_poss.replace = false; + } + + if (pi->second.replace) + throw skip_configuration (move (ed), pi->second); + else + throw skip_configuration (); + } + + // The other clusters check is a bit more complicated: if the + // other cluster (with the earlier position) is not yet + // negotiated, then we skip. Otherwise, we have to add an entry + // to postponed_poss and backtrack. + // + bool skip (false); + for (const postponed_configuration& cfg: postponed_cfgs) + { + // Skip the current cluster. + // + if (&cfg == pcfg) + continue; + + if (const pair* p = + cfg.existing_dependent_position (pk)) + { + size_t ei (p->first); // Other position. + + if (!cfg.negotiated) + { + if (ei < di) + { + l5 ([&]{trace << "cannot re-evaluate dependent " + << pk << " to dependency index " << di + << " due to earlier dependency index " + << ei << " in " << cfg << ", skipping " + << *pcfg;}); + + skip = true; + } + } + else + { + // If this were not the case, then this dependent wouldn't + // have been considered as an existing by + // query_existing_dependents() since as it is (being) + // negotiated then it is already re-evaluated and so is + // being built (see the verify lambda above). + // + assert (ei > di); + + // Feels like there cannot be an earlier position. + // + postponed_position pp (ed.dependency_position, + false /* replace */); + + auto p (postponed_poss.emplace (pk, pp)); + if (!p.second) + { + assert (p.first->second > pp); + p.first->second = pp; + } + + l5 ([&]{trace << "cannot re-evaluate dependent " + << pk << " to dependency index " << di + << " due to greater dependency " + << "index " << ei << " in " << cfg + << ", throwing postpone_position";}); + + throw postpone_position (); + } + } + } + + if (skip) + throw skip_configuration (); + + // Finally, re-evaluate the dependent. + // + packages& ds (ed.dependencies); + + pair, + lazy_shared_ptr> rp ( + find_available_fragment (o, pk.db, ed.selected)); + + build_package p { + build_package::build, + pk.db, + move (ed.selected), + move (rp.first), + move (rp.second), + nullopt, // Dependencies. + nullopt, // Dependencies alternatives. + nullopt, // Package skeleton. + nullopt, // Postponed dependency alternatives. + false, // Recursive collection. + nullopt, // Hold package. + nullopt, // Hold version. + {}, // Constraints. + false, // System. + false, // Keep output directory. + false, // Disfigure (from-scratch reconf). + false, // Configure-only. + nullopt, // Checkout root. + false, // Checkout purge. + strings (), // Configuration variables. + set ( // Required by (dependency). + ds.begin (), ds.end ()), + false, // Required by dependents. + build_package::adjust_reconfigure | + build_package::build_reevaluate}; + + // Note: not recursive. + // + collect_build (o, + move (p), + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + postponed_cfgs); + + build_package* b (entered_build (pk)); + assert (b != nullptr); + + // Re-evaluate up to the earliest position. + // + assert (ed.dependency_position.first != 0); + + build_package_refs dep_chain; + collect_build_prerequisites (o, + *b, + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + dep_chain, + &postponed_repo, + &postponed_alts, + numeric_limits::max (), + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts, + ed.dependency_position); + + ed.reevaluated = true; + + if (pi != postponed_poss.end ()) + { + // Otherwise we should have thrown skip_configuration above. + // + assert (di <= pi->second.first); + + pi->second.reevaluated = true; + } + } + } + } + } + + l5 ([&]{trace << "cfg-negotiate begin " << *pcfg;}); + + // Negotiate the configuration. + // + // The overall plan is as follows: continue refining the configuration + // until there are no more changes by giving each dependent a chance to + // make further adjustments. + // + for (auto b (pcfg->dependents.begin ()), + i (b), + e (pcfg->dependents.end ()); i != e; ) + { + // Resolve package skeletons for the dependent and its dependencies. + // + // For the dependent, the skeleton should be already there (since we + // should have started recursively collecting it). For a dependency, + // it should not already be there (since we haven't yet started + // recursively collecting it). But we could be re-resolving the same + // dependency multiple times. + // + package_skeleton* dept; + { + build_package* b (entered_build (i->first)); + assert (b != nullptr && b->skeleton); + dept = &*b->skeleton; + } + + pair pos; + small_vector, 1> depcs; + bool has_alt; + { + // A non-negotiated cluster must only have one depends position for + // each dependent. + // + assert (i->second.dependencies.size () == 1); + + const postponed_configuration::dependency& ds ( + i->second.dependencies.front ()); + + pos = ds.position; + + // Note that an existing dependent which initially doesn't have the + // has_alternative flag present should obtain it as a part of + // re-evaluation at this time. + // + assert (ds.has_alternative); + + has_alt = *ds.has_alternative; + + depcs.reserve (ds.size ()); + for (const package_key& pk: ds) + { + build_package* b (entered_build (pk)); + assert (b != nullptr); + + depcs.push_back (b->skeleton + ? *b->skeleton + : b->init_skeleton (o /* options */)); + } + } + + optional changed ( + negotiate_configuration ( + pcfg->dependency_configurations, *dept, pos, depcs, has_alt)); + + // If the dependency alternative configuration cannot be negotiated + // for this dependent, then add an entry to unacceptable_alts and + // throw unaccept_alternative to recollect from scratch. + // + if (!changed) + { + assert (dept->available != nullptr); // Can't be system. + + const package_key& p (dept->package); + const version& v (dept->available->version); + + unacceptable_alts.emplace (p, v, pos); + + l5 ([&]{trace << "unable to cfg-negotiate dependency alternative " + << pos.first << ',' << pos.second << " for " + << "dependent " << package_string (p.name, v) + << p.db << ", throwing unaccept_alternative";}); + + throw unaccept_alternative (); + } + else if (*changed) + { + if (i != b) + { + i = b; // Restart from the beginning. + continue; + } + } + + ++i; + } + + // Being negotiated (so can only be up-negotiated). + // + pcfg->negotiated = false; + + // Note that we can be adding new packages to the being negotiated + // cluster by calling collect_build_prerequisites() for its dependencies + // and dependents. Thus, we need to stash the current list of + // dependencies and dependents and iterate over them. + // + // Note that whomever is adding new packages is expected to process them + // (they may also process existing packages, which we are prepared to + // ignore). + // + packages dependencies (pcfg->dependencies); + + packages dependents; + dependents.reserve (pcfg->dependents.size ()); + + for (const auto& p: pcfg->dependents) + dependents.push_back (p.first); + + // Process dependencies recursively with this config. + // + // Note that there could be inter-dependecies between these packages, + // which means the configuration can only be up-negotiated. + // + l5 ([&]{trace << "recursively collect cfg-negotiated dependencies";}); + + for (const package_key& p: dependencies) + { + build_package* b (entered_build (p)); + assert (b != nullptr); + + // Reconfigure the configured dependencies. + // + // Note that potentially this can be an overkill if the dependency + // configuration doesn't really change. Later we can implement some + // precise detection for that using configuration checksum or similar. + // + // Also note that for configured dependents which belong to the + // configuration cluster this flag is already set (see above). + // + if (b->selected != nullptr && + b->selected->state == package_state::configured) + b->flags |= build_package::adjust_reconfigure; + + // Skip the dependencies which are already collected recursively. + // + if (!b->recursive_collection) + { + // Verify and set the dependent configuration for this dependency. + // + // Note: see similar code for the up-negotiation case. + // + { + assert (b->skeleton); // Should have been init'ed above. + + const package_configuration& pc ( + pcfg->dependency_configurations[p]); + + // Skip the verification if this is a system package without + // skeleton info. + // + pair pr (b->skeleton->available != nullptr + ? b->skeleton->verify_sensible (pc) + : make_pair (true, string ())); + + if (!pr.first) + { + // Note that the diagnostics from the dependency will most + // likely be in the "error ..." form (potentially with + // additional info lines) and by printing it with a two-space + // indentation we make it "fit" into our diag record. + // + diag_record dr (fail); + dr << "unable to negotiate sensible configuration for " + << "dependency " << p << '\n' + << " " << pr.second; + + dr << info << "negotiated configuration:\n"; + pc.print (dr, " "); // Note 4 spaces since in nested info. + } + + b->skeleton->dependent_config (pc); + } + + build_package_refs dep_chain; + collect_build_prerequisites (o, + *b, + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + dep_chain, + &postponed_repo, + &postponed_alts, + 0 /* max_alt_index */, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + } + else + l5 ([&]{trace << "dependency " << b->available_name_version_db () + << " is already (being) recursively collected, " + << "skipping";}); + } + + // Continue processing dependents with this config. + // + l5 ([&]{trace << "recursively collect cfg-negotiated dependents";}); + + for (const auto& p: dependents) + { + // Select the dependency alternative for which configuration has been + // negotiated and collect this dependent starting from the next + // depends value. + // + build_package* b (entered_build (p)); + + // We should have been started recursively collecting the dependent + // and it should have been postponed. + // + assert (b != nullptr && + b->available != nullptr && + b->dependencies && + b->skeleton && + b->postponed_dependency_alternatives); + + // Select the dependency alternative (evaluate reflect if present, + // etc) and position to the next depends value (see + // collect_build_prerequisites() for details). + // + { + const bpkg::dependencies& deps (b->available->dependencies); + bpkg::dependencies& sdeps (*b->dependencies); + vector& salts (*b->alternatives); + + size_t di (sdeps.size ()); + + // Skip the dependent if it has been already collected as some + // package's dependency or some such. + // + if (di == deps.size ()) + { + l5 ([&]{trace << "dependent " << b->available_name_version_db () + << " is already recursively collected, skipping";}); + + continue; + } + + l5 ([&]{trace << "select cfg-negotiated dependency alternative " + << "for dependent " + << b->available_name_version_db ();}); + + // Find the postponed dependency alternative. + // + auto i (pcfg->dependents.find (p)); + + assert (i != pcfg->dependents.end () && + i->second.dependencies.size () == 1); + + pair dp (i->second.dependencies[0].position); + assert (dp.first == sdeps.size () + 1); + + build_package::dependency_alternatives_refs pdas ( + move (*b->postponed_dependency_alternatives)); + + b->postponed_dependency_alternatives = nullopt; + + auto j (find_if (pdas.begin (), pdas.end (), + [&dp] (const auto& da) + { + return da.second + 1 == dp.second; + })); + + assert (j != pdas.end ()); + + const dependency_alternative& da (j->first); + size_t dai (j->second); + + // Select the dependency alternative and position to the next + // depends value. + // + const dependency_alternatives_ex& das (deps[di]); + dependency_alternatives_ex sdas (das.buildtime, das.comment); + + sdas.emplace_back (nullopt /* enable */, + nullopt /* reflect */, + da.prefer, + da.accept, + da.require, + da /* dependencies */); + + sdeps.push_back (move (sdas)); + salts.push_back (dai); + + // Evaluate reflect, if present. + // + if (da.reflect) + b->skeleton->evaluate_reflect (*da.reflect, make_pair (di, dai)); + } + + // Continue recursively collecting the dependent. + // + build_package_refs dep_chain; + + collect_build_prerequisites (o, + *b, + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + dep_chain, + &postponed_repo, + &postponed_alts, + 0 /* max_alt_index */, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + } + + // Negotiated (so can only be rolled back). + // + pcfg->negotiated = true; + + l5 ([&]{trace << "cfg-negotiate end " << *pcfg;}); + + // Fall through (to start another iteration of the below loop). + } + + // Try collecting postponed packages for as long as we are making + // progress. + // + vector spas; // Reuse. + + for (bool prog (!postponed_repo.empty () || + !postponed_cfgs.negotiated () || + !postponed_alts.empty () || + postponed_deps.has_bogus ()); + prog; ) + { + postponed_packages prs; + postponed_packages pas; + + // Try to collect the repository-related postponments first. + // + for (build_package* p: postponed_repo) + { + l5 ([&]{trace << "collect rep-postponed " + << p->available_name_version_db ();}); + + build_package_refs dep_chain; + + collect_build_prerequisites (o, + *p, + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + dep_chain, + &prs, + &pas, + 0 /* max_alt_index */, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + } + + // Save the potential new dependency alternative-related postponements. + // + postponed_alts.insert (pas.begin (), pas.end ()); + + prog = (prs != postponed_repo); + + if (prog) + { + postponed_repo.swap (prs); + continue; + } + + // Now, as there is no more progress made in collecting repository- + // related postponements, collect the dependency configuration-related + // postponements. + // + // Note that we do it before alternatives since configurations we do + // perfectly (via backtracking) while alternatives -- heuristically. + // + // Note that since the potential snapshot restore replaces all the list + // entries we cannot iterate using the iterator here. Also note that the + // list size may change during iterating. + // + for (size_t ci (0); ci != postponed_cfgs.size (); ++ci) + { + postponed_configuration* pc (&postponed_cfgs[ci]); + + // Find the next configuration to try to negotiate, skipping the + // already negotiated ones. + // + if (pc->negotiated) + continue; + + size_t pcd (depth + 1); + pc->depth = pcd; + + // Either return or retry the same cluster or skip this cluster and + // proceed to the next one. + // + for (;;) + { + // First assume we can negotiate this configuration rolling back if + // this doesn't pan out. + // + snapshot s (*this, + postponed_repo, + postponed_alts, + postponed_deps, + postponed_cfgs); + + try + { + collect_build_postponed (o, + replaced_vers, + postponed_repo, + postponed_alts, + postponed_deps, + postponed_cfgs, + postponed_cfgs_history, + postponed_poss, + unacceptable_alts, + fdb, + rpt_depts, + apc, + pc); + + // If collect() returns (instead of throwing), this means it + // processed everything that was postponed. + // + assert (postponed_repo.empty () && + postponed_cfgs.negotiated () && + postponed_alts.empty () && + !postponed_deps.has_bogus ()); + + l5 ([&]{trace << "end" << trace_suffix;}); + + return; + } + catch (skip_configuration& e) + { + // Restore the state from snapshot. + // + // Note: postponed_cfgs is re-assigned. + // + s.restore (*this, + postponed_repo, + postponed_alts, + postponed_deps, + postponed_cfgs); + + pc = &postponed_cfgs[ci]; + + // Note that in this case we keep the accumulated configuration, + // if any. + + pc->depth = 0; + + // If requested, "replace" the "later" dependent-dependency + // cluster with an earlier. + // + if (e.dependent) + { + existing_dependent& ed (*e.dependent); + pair pos (e.new_position); + + const build_package* bp ( + replace_existing_dependent_dependency ( + trace, + o, + ed, // Note: modified. + pos, + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + postponed_cfgs)); + + postponed_cfgs.add (package_key (ed.db, ed.selected->name), + pos, + package_key (bp->db, bp->selected->name)); + } + + l5 ([&]{trace << "postpone cfg-negotiation of " << *pc;}); + + break; + } + catch (const retry_configuration& e) + { + // If this is not "our problem", then keep looking. + // + if (e.depth != pcd) + throw; + + package_configurations cfgs ( + move (pc->dependency_configurations)); + + // Restore the state from snapshot. + // + // Note: postponed_cfgs is re-assigned. + // + s.restore (*this, + postponed_repo, + postponed_alts, + postponed_deps, + postponed_cfgs); + + pc = &postponed_cfgs[ci]; + + l5 ([&]{trace << "cfg-negotiation of " << *pc << " failed due " + << "to dependent " << e.dependent << ", refining " + << "configuration";}); + + // Copy over the configuration for further refinement. + // + // Note that there is also a possibility of ending up with "bogus" + // configuration variables that were set by a dependent during + // up-negotiation but, due to changes to the overall + // configuration, such a dependent were never re-visited. + // + // The way we are going to deal with this is by detecting such + // bogus variables based on the confirmed flag, cleaning them out, + // and doing another retry. Here we clear the confirmed flag and + // the detection happens in collect_build_postponed() after we + // have processed everything postponed (since that's the only time + // we can be certain there could no longer be a re-visit). + // + for (package_configuration& cfg: cfgs) + for (config_variable_value& v: cfg) + if (v.dependent) + v.confirmed = false; + + pc->dependency_configurations = move (cfgs); + } + catch (merge_configuration& e) + { + // If this is not "our problem", then keep looking. + // + if (e.depth != pcd) + throw; + + postponed_configuration shadow (move (*pc)); + + // Restore the state from snapshot. + // + // Note: postponed_cfgs is re-assigned. + // + s.restore (*this, + postponed_repo, + postponed_alts, + postponed_deps, + postponed_cfgs); + + pc = &postponed_cfgs[ci]; + + assert (!pc->negotiated); + + // Drop any accumulated configuration (which could be carried + // over from retry_configuration logic). + // + pc->dependency_configurations.clear (); + + l5 ([&]{trace << "cfg-negotiation of " << *pc << " failed due " + << "to non-negotiated clusters, force-merging " + << "based on shadow cluster " << shadow;}); + + // Pre-merge into this cluster those non-negotiated clusters which + // were merged into the shadow cluster. + // + for (size_t id: shadow.merged_ids) + { + postponed_configuration* c (postponed_cfgs.find (id)); + + if (c != nullptr) + { + // Otherwise we would be handling the exception in the higher + // stack frame. + // + assert (!c->negotiated); + + l5 ([&]{trace << "force-merge " << *c << " into " << *pc;}); + + pc->merge (move (*c)); + + // Mark configuration as the one being merged from for + // subsequent erasing from the list. + // + c->dependencies.clear (); + } + } + + // Erase clusters which we have merged from. Also re-translate the + // current cluster address into index which may change as a result + // of the merge. + // + auto i (postponed_cfgs.begin ()); + auto j (postponed_cfgs.before_begin ()); // Precedes iterator i. + + for (size_t k (0); i != postponed_cfgs.end (); ) + { + if (!i->dependencies.empty ()) + { + if (&*i == pc) + ci = k; + + ++i; + ++j; + ++k; + } + else + i = postponed_cfgs.erase_after (j); + } + + pc->set_shadow_cluster (move (shadow)); + } + } + } + + // Note that we only get here if we didn't make any progress on the + // previous loop (the only "progress" path ends with return). + + // Now, try to collect the dependency alternative-related + // postponements. + // + if (!postponed_alts.empty ()) + { + // Sort the postponments in the unprocessed dependencies count + // descending order. + // + // The idea here is to preferably handle those postponed packages + // first, which have a higher probability to affect the dependency + // alternative selection for other packages. + // + spas.assign (postponed_alts.begin (), postponed_alts.end ()); + + std::sort (spas.begin (), spas.end (), + [] (build_package* x, build_package* y) + { + size_t xt (x->available->dependencies.size () - + x->dependencies->size ()); + + size_t yt (y->available->dependencies.size () - + y->dependencies->size ()); + + if (xt != yt) + return xt > yt ? -1 : 1; + + // Also factor the package name and configuration path + // into the ordering to achieve a stable result. + // + int r (x->name ().compare (y->name ())); + return r != 0 + ? r + : x->db.get ().config.compare (y->db.get ().config); + }); + + // Calculate the maximum number of the enabled dependency + // alternatives. + // + size_t max_enabled_count (0); + + for (build_package* p: spas) + { + assert (p->postponed_dependency_alternatives); + + size_t n (p->postponed_dependency_alternatives->size ()); + + if (max_enabled_count < n) + max_enabled_count = n; + } + + assert (max_enabled_count != 0); // Wouldn't be here otherwise. + + // Try to select a dependency alternative with the lowest index, + // preferring postponed packages with the longer tail of unprocessed + // dependencies (see above for the reasoning). + // + for (size_t i (1); i <= max_enabled_count && !prog; ++i) + { + for (build_package* p: spas) + { + prs.clear (); + pas.clear (); + + size_t ndep (p->dependencies->size ()); + + build_package_refs dep_chain; + + l5 ([&]{trace << "index " << i << " collect alt-postponed " + << p->available_name_version_db ();}); + + collect_build_prerequisites (o, + *p, + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + dep_chain, + &prs, + &pas, + i, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + + prog = (pas.find (p) == pas.end () || + ndep != p->dependencies->size ()); + + // Save the potential new postponements. + // + if (prog) + { + postponed_alts.erase (p); + postponed_alts.insert (pas.begin (), pas.end ()); + } + + size_t npr (postponed_repo.size ()); + postponed_repo.insert (prs.begin (), prs.end ()); + + // Note that not collecting any alternative-relative postponements + // but producing new repository-related postponements is progress + // nevertheless. + // + // Note that we don't need to check for new configuration- related + // postponements here since if they are present, then this package + // wouldn't be in pas and so prog would be true (see above for + // details). + // + if (!prog) + prog = (npr != postponed_repo.size ()); + + if (prog) + break; + } + } + + if (prog) + continue; + } + + assert (!prog); + + // If we still have any non-negotiated clusters and non-replace + // postponed positions, then it's possible one of them is the cross- + // dependent pathological case where we will never hit it unless we + // force the re-evaluation to earlier position (similar to the + // single-dependent case, which we handle accurately). For example: + // + // tex: depends: libbar(c) + // depends: libfoo(c) + // + // tix: depends: libbar(c) + // depends: tex(c) + // + // Here tex and tix are existing dependent and we are upgrading tex. + // + // While it would be ideal to handle such cases accurately, it's not + // trivial. So for now we resort to the following heuristics: when left + // with no other option, we treat the first encountered non- replace + // position as replace and see if that helps move things forward. + // + if (!postponed_cfgs.negotiated () && + find_if (postponed_poss.begin (), postponed_poss.end (), + [] (const auto& v) {return !v.second.replace;}) != + postponed_poss.end () && + !postponed_poss.replace) + { + l5 ([&]{trace << "non-negotiated clusters left and non-replace " + << "postponed positions are present, overriding first " + << "encountered non-replace position to replace";}); + + postponed_poss.replace = true; + prog = true; + continue; // Go back to negotiating skipped cluster. + } + + // Finally, erase the bogus postponements and re-collect from scratch, + // if any (see postponed_dependencies for details). + // + // Note that we used to re-collect such postponements in-place but + // re-doing from scratch feels more correct (i.e., we may end up doing + // it earlier which will affect dependency alternatives). + // + postponed_deps.cancel_bogus (trace, false /* initial_collection */); + } + + // Check if any negotiatiated configurations ended up with any bogus + // variables (see retry_configuration catch block for background). + // + // Note that we could potentially end up yo-yo'ing: we remove a bogus and + // that causes the original dependent to get re-visited which in turn + // re-introduces the bogus. In other words, one of the bogus variables + // which we have removed are actually the cause of no longer needing the + // dependent that introduced it. Feels like the correct outcome of this + // should be keeping the bogus variable that triggered yo-yo'ing. Of + // course, there could be some that we should keep and some that we should + // drop and figuring this out would require retrying all possible + // combinations. An alternative solution would be to detect yo-yo'ing, + // print the bogus variables involved, and ask the user to choose (with an + // override) which ones to keep. Let's go with this for now. + // + { + // On the first pass see if we have anything bogus. + // + bool bogus (false); + for (postponed_configuration& pcfg: postponed_cfgs) + { + if (pcfg.negotiated && *pcfg.negotiated) // Negotiated. + { + for (package_configuration& cfg: pcfg.dependency_configurations) + { + for (config_variable_value& v: cfg) + { + if (v.dependent && !v.confirmed) + { + bogus = true; + break; + } + } + if (bogus) break; + } + if (bogus) break; + } + } + + if (bogus) + { + // On the second pass calculate the checksum of all the negotiated + // clusters. + // + sha256 cs; + for (postponed_configuration& pcfg: postponed_cfgs) + { + if (pcfg.negotiated && *pcfg.negotiated) + { + for (package_configuration& cfg: pcfg.dependency_configurations) + { + for (config_variable_value& v: cfg) + { + if (v.dependent) + to_checksum (cs, v); + } + } + } + } + + bool cycle; + { + string s (cs.string ()); + if (find (postponed_cfgs_history.begin (), + postponed_cfgs_history.end (), + s) == postponed_cfgs_history.end ()) + { + postponed_cfgs_history.push_back (move (s)); + cycle = false; + } + else + cycle = true; + } + + // On the third pass we either retry or diagnose. + // + diag_record dr; + if (cycle) + { + dr << + fail << "unable to remove bogus configuration values without " + << "causing configuration refinement cycle" << + info << "consider manually specifying one or more of the " + << "following variables as user configuration"; + } + + for (postponed_configuration& pcfg: postponed_cfgs) + { + optional dept; // Bogus dependent. + + if (pcfg.negotiated && *pcfg.negotiated) + { + for (package_configuration& cfg: pcfg.dependency_configurations) + { + // Note that the entire dependency configuration may end up + // being "bogus" (i.e., it does not contain any configuration + // variables with a confirmed dependent). But that will be + // handled naturally: we will either no longer have this + // dependency in the cluster and thus never call its skeleton's + // dependent_config() or this call will be no-op since it won't + // find any dependent variables. + // + for (config_variable_value& v: cfg) + { + if (v.dependent && !v.confirmed) + { + if (!dept) + dept = move (v.dependent); + + if (cycle) + dr << "\n " << v.serialize_cmdline (); + else + v.undefine (); + } + } + } + + if (dept) + { + if (cycle) + break; + else + throw retry_configuration {pcfg.depth, move (*dept)}; + } + } + + if (dept) + break; + } + } + } + + // If any postponed_{repo,alts} builds remained, then perform the + // diagnostics run. Naturally we shouldn't have any postponed_cfgs without + // one of the former. + // + if (!postponed_repo.empty ()) + { + build_package_refs dep_chain; + + collect_build_prerequisites (o, + **postponed_repo.begin (), + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + dep_chain, + nullptr, + nullptr, + 0, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + + assert (false); // Can't be here. + } + + if (!postponed_alts.empty ()) + { + build_package_refs dep_chain; + + collect_build_prerequisites (o, + **postponed_alts.begin (), + fdb, + rpt_depts, + apc, + false /* initial_collection */, + replaced_vers, + dep_chain, + nullptr, + nullptr, + 0, + postponed_deps, + postponed_cfgs, + postponed_poss, + unacceptable_alts); + + assert (false); // Can't be here. + } + + // While the assumption is that we shouldn't leave any non-negotiated + // clusters, we can potentially miss some corner cases in the above "skip + // configuration" logic. Let's thus trace the non-negotiated clusters + // before the assertion. + // +#ifndef NDEBUG + for (const postponed_configuration& cfg: postponed_cfgs) + { + if (!cfg.negotiated || !*cfg.negotiated) + trace << "unexpected non-negotiated cluster " << cfg; + } + + assert (postponed_cfgs.negotiated ()); +#endif + + l5 ([&]{trace << "end" << trace_suffix;}); + } + + build_packages::iterator build_packages:: + order (database& db, + const package_name& name, + optional buildtime, + const function& fdb, + bool reorder) + { + package_refs chain; + return order (db, name, buildtime, chain, fdb, reorder); + } + + void build_packages:: + collect_order_dependents (const repointed_dependents& rpt_depts) + { + // For each package on the list we want to insert all its dependents + // before it so that they get configured after the package on which they + // depend is configured (remember, our build order is reverse, with the + // last package being built first). This applies to both packages that are + // already on the list as well as the ones that we add, recursively. + // + for (auto i (begin ()); i != end (); ++i) + { + const build_package& p (*i); + + // Prune if this is not a configured package being up/down-graded + // or reconfigured. + // + assert (p.action); + + // Dropped package may have no dependents. + // + if (*p.action != build_package::drop && p.reconfigure ()) + collect_order_dependents (i, rpt_depts); + } + } + + void build_packages:: + collect_order_dependents (iterator pos, + const repointed_dependents& rpt_depts) + { + tracer trace ("collect_order_dependents"); + + assert (pos != end ()); + + build_package& p (*pos); + + database& pdb (p.db); + const shared_ptr& sp (p.selected); + + const package_name& n (sp->name); + + // See if we are up/downgrading this package. In particular, the available + // package could be NULL meaning we are just adjusting. + // + int ud (p.available != nullptr + ? sp->version.compare (p.available_version ()) + : 0); + + for (database& ddb: pdb.dependent_configs ()) + { + for (auto& pd: query_dependents_cache (ddb, n, pdb)) + { + package_name& dn (pd.name); + auto i (map_.find (ddb, dn)); + + // Make sure the up/downgraded package still satisfies this + // dependent. But first "prune" if the dependent is being dropped or + // this is a replaced prerequisite of the repointed dependent. + // + // Note that the repointed dependents are always collected and have + // all their collected prerequisites ordered (including new and old + // ones). See collect_build_prerequisites() and order() for details. + // + bool check (ud != 0 && pd.constraint); + + if (i != map_.end () && i->second.position != end ()) + { + build_package& dp (i->second.package); + + // Skip the droped dependent. + // + if (dp.action && *dp.action == build_package::drop) + continue; + + repointed_dependents::const_iterator j ( + rpt_depts.find (package_key {ddb, dn})); + + if (j != rpt_depts.end ()) + { + const map& prereqs_flags (j->second); + + auto k (prereqs_flags.find (package_key {pdb, n})); + + if (k != prereqs_flags.end () && !k->second) + continue; + } + + // There is one tricky aspect: the dependent could be in the process + // of being reconfigured or up/downgraded as well. In this case all + // we need to do is detect this situation and skip the test since + // all the (new) constraints of this package have been satisfied in + // collect_build(). + // + if (check) + check = !dp.dependencies; + } + + if (check) + { + const version& av (p.available_version ()); + const version_constraint& c (*pd.constraint); + + if (!satisfies (av, c)) + { + diag_record dr (fail); + + dr << "unable to " << (ud < 0 ? "up" : "down") << "grade " + << "package " << *sp << pdb << " to "; + + // Print both (old and new) package names in full if the system + // attribution changes. + // + if (p.system != sp->system ()) + dr << p.available_name_version (); + else + dr << av; // Can't be the wildcard otherwise would satisfy. + + dr << info << "because package " << dn << ddb << " depends on (" + << n << " " << c << ")"; + + string rb; + if (!p.user_selection ()) + { + for (const package_key& pk: p.required_by) + rb += (rb.empty () ? " " : ", ") + pk.string (); + } + + if (!rb.empty ()) + dr << info << "package " << p.available_name_version () + << " required by" << rb; + + dr << info << "explicitly request up/downgrade of package " + << dn; + + dr << info << "or explicitly specify package " << n + << " version to manually satisfy these constraints"; + } + + // Add this contraint to the list for completeness. + // + p.constraints.emplace_back (ddb, dn.string (), c); + } + + auto adjustment = [&dn, &ddb, &n, &pdb] () -> build_package + { + shared_ptr dsp (ddb.load (dn)); + + // A system package cannot be a dependent. + // + assert (!dsp->system ()); + + return build_package { + build_package::adjust, + ddb, + move (dsp), + nullptr, // No available pkg/repo fragment. + nullptr, + nullopt, // Dependencies. + nullopt, // Dependencies alternatives. + nullopt, // Package skeleton. + nullopt, // Postponed dependency alternatives. + false, // Recursive collection. + nullopt, // Hold package. + nullopt, // Hold version. + {}, // Constraints. + false, // System. + false, // Keep output directory. + false, // Disfigure (from-scratch reconf). + false, // Configure-only. + nullopt, // Checkout root. + false, // Checkout purge. + strings (), // Configuration variables. + {package_key {pdb, n}}, // Required by (dependency). + false, // Required by dependents. + build_package::adjust_reconfigure}; + }; + + // We can have three cases here: the package is already on the list, + // the package is in the map (but not on the list) and it is in + // neither. + // + // If the existing entry is pre-entered, is an adjustment, or is a + // build that is not supposed to be built (not in the list), then we + // merge it into the new adjustment entry. Otherwise (is a build in + // the list), we just add the reconfigure adjustment flag to it. + // + if (i != map_.end ()) + { + build_package& dp (i->second.package); + iterator& dpos (i->second.position); + + if (!dp.action || // Pre-entered. + *dp.action != build_package::build || // Non-build. + dpos == end ()) // Build not in the list. + { + build_package bp (adjustment ()); + bp.merge (move (dp)); + dp = move (bp); + } + else // Build in the list. + dp.flags |= build_package::adjust_reconfigure; + + // It may happen that the dependent is already in the list but is + // not properly ordered against its dependencies that get into the + // list via another dependency path. Thus, we check if the dependent + // is to the right of its dependency and, if that's the case, + // reinsert it in front of the dependency. + // + if (dpos != end ()) + { + for (auto i (pos); i != end (); ++i) + { + if (i == dpos) + { + erase (dpos); + dpos = insert (pos, dp); + break; + } + } + } + else + dpos = insert (pos, dp); + } + else + { + // Don't move dn since it is used by adjustment(). + // + i = map_.emplace (package_key {ddb, dn}, + data_type {end (), adjustment ()}).first; + + i->second.position = insert (pos, i->second.package); + } + + // Recursively collect our own dependents inserting them before us. + // + // Note that we cannot end up with an infinite recursion for + // configured packages due to a dependency cycle (see order() for + // details). + // + collect_order_dependents (i->second.position, rpt_depts); + } + } + } + + void build_packages:: + clear () + { + build_package_list::clear (); + map_.clear (); + } + + void build_packages:: + clear_order () + { + build_package_list::clear (); + + for (auto& p: map_) + p.second.position = end (); + } + + void build_packages:: + verify_ordering () const + { + for (const auto& b: map_) + { + const build_package& bp (b.second.package); + + auto i (find_if (begin (), end (), + [&bp] (const build_package& p) {return &p == &bp;})); + + // List ordering must properly be reflected in the tree entries. + // + assert (i == b.second.position); + + // Pre-entered builds must never be ordered and the real build actions + // (builds, adjustments, etc) must all be ordered. + // + // Note that the later was not the case until we've implemented + // re-collection from scratch after the package version replacement (see + // replaced_versions for details). Before that the whole dependency + // trees from the being replaced dependent stayed in the map. + // + assert (bp.action.has_value () == (i != end ())); + } + } + + vector build_packages:: + query_existing_dependents ( + tracer& trace, + database& db, + const package_name& name, + const replaced_versions& replaced_vers, + const repointed_dependents& rpt_depts, + const function& vdb) + { + vector r; + + lazy_shared_ptr sp (db, name); + + for (database& ddb: db.dependent_configs ()) + { + for (auto& pd: query_dependents (ddb, name, db)) + { + shared_ptr dsp ( + ddb.load (pd.name)); + + auto i (dsp->prerequisites.find (sp)); + assert (i != dsp->prerequisites.end ()); + + const auto& pos (i->second.config_position); + + if (pos.first != 0) // Has config clause? + { + package_key pk (ddb, pd.name); + + if (rpt_depts.find (pk) != rpt_depts.end ()) + { + l5 ([&]{trace << "skip repointed existing dependent " << pk + << " of dependency " << name << db;}); + continue; + } + + // Ignore dependent which is already being built or dropped. + // + const build_package* p (entered_build (pk)); + + if (p != nullptr && p->action) + { + bool build; + if (((build = *p->action == build_package::build) && + (p->system || p->recollect_recursively (rpt_depts))) || + *p->action == build_package::drop) + { + if (!build || !vdb || !vdb (pk, pos)) + { + l5 ([&]{trace << "skip being " + << (build ? "built" : "dropped") + << " existing dependent " << pk + << " of dependency " << name << db;}); + continue; + } + } + } + + // Ignore dependent which is expected to be built or dropped. + // + auto vi (replaced_vers.find (pk)); + if (vi != replaced_vers.end () && !vi->second.replaced) + { + bool build (vi->second.available != nullptr); + + l5 ([&]{trace << "skip expected to be " + << (build ? "built" : "dropped") + << " existing dependent " << pk + << " of dependency " << name << db;}); + + continue; + } + + r.push_back (existing_dependent {ddb, move (dsp), pos}); + } + } + } + + return r; + } + + const build_package* build_packages:: + replace_existing_dependent_dependency ( + tracer& trace, + const pkg_build_options& o, + existing_dependent& ed, + pair pos, + const function& fdb, + const repointed_dependents& rpt_depts, + const function& apc, + bool initial_collection, + replaced_versions& replaced_vers, + postponed_configurations& postponed_cfgs) + { + // The repointed dependent cannot be returned by + // query_existing_dependents(). Note that the repointed dependent + // references both old and new prerequisites. + // + assert (rpt_depts.find (package_key (ed.db, ed.selected->name)) == + rpt_depts.end ()); + + shared_ptr dsp; + database* pdb (nullptr); + const version_constraint* vc (nullptr); + + // Find the dependency for this earlier dependency position. We know it + // must be there since it's with configuration. + // + for (const auto& p: ed.selected->prerequisites) + { + if (p.second.config_position == pos) + { + pdb = &p.first.database (); + + dsp = p.first.load (); + + l5 ([&]{trace << "replace dependency at index " + << ed.dependency_position.first + << " of existing dependent " << *ed.selected + << ed.db << " with dependency " << *dsp + << *pdb << " at index " << pos.first;}); + + if (p.second.constraint) + vc = &*p.second.constraint; + } + } + + assert (dsp != nullptr); + + package_key pk (*pdb, dsp->name); + + // Adjust the existing dependent entry. + // + ed.dependency_position = pos; + + // Collect the package build for this dependency. + // + pair, + lazy_shared_ptr> rp ( + find_available_fragment (o, pk.db, dsp)); + + bool system (dsp->system ()); + + package_key dpk (ed.db, ed.selected->name); + + build_package p { + build_package::build, + pk.db, + move (dsp), + move (rp.first), + move (rp.second), + nullopt, // Dependencies. + nullopt, // Dependencies alternatives. + nullopt, // Package skeleton. + nullopt, // Postponed dependency alternatives. + false, // Recursive collection. + nullopt, // Hold package. + nullopt, // Hold version. + {}, // Constraints. + system, // System. + false, // Keep output directory. + false, // Disfigure (from-scratch reconf). + false, // Configure-only. + nullopt, // Checkout root. + false, // Checkout purge. + strings (), // Configuration variables. + {dpk}, // Required by (dependent). + true, // Required by dependents. + build_package::adjust_reconfigure}; + + if (vc != nullptr) + p.constraints.emplace_back (dpk.db, dpk.name.string (), *vc); + + // Note: not recursive. + // + collect_build (o, + move (p), + fdb, + rpt_depts, + apc, + initial_collection, + replaced_vers, + postponed_cfgs); + + return entered_build (pk); + } + + build_packages::iterator build_packages:: + order (database& db, + const package_name& name, + optional buildtime, + package_refs& chain, + const function& fdb, + bool reorder) + { + package_map::iterator mi; + + if (buildtime) + { + database* ddb (fdb (db, name, *buildtime)); + + mi = ddb != nullptr + ? map_.find (*ddb, name) + : map_.find_dependency (db, name, *buildtime); + } + else + mi = map_.find (db, name); + + // Every package that we order should have already been collected. + // + assert (mi != map_.end ()); + + build_package& p (mi->second.package); + + assert (p.action); // Can't order just a pre-entered package. + + database& pdb (p.db); + + // Make sure there is no dependency cycle. + // + package_ref cp {pdb, name}; + { + auto i (find (chain.begin (), chain.end (), cp)); + + if (i != chain.end ()) + { + diag_record dr (fail); + dr << "dependency cycle detected involving package " << name << pdb; + + auto nv = [this] (const package_ref& cp) + { + auto mi (map_.find (cp.db, cp.name)); + assert (mi != map_.end ()); + + build_package& p (mi->second.package); + + assert (p.action); // See above. + + // We cannot end up with a dependency cycle for actions other than + // build since these packages are configured and we would fail on a + // previous run while building them. + // + assert (p.available != nullptr); + + return p.available_name_version_db (); + }; + + // Note: push_back() can invalidate the iterator. + // + size_t j (i - chain.begin ()); + + for (chain.push_back (cp); j != chain.size () - 1; ++j) + dr << info << nv (chain[j]) << " depends on " << nv (chain[j + 1]); + } + } + + // If this package is already in the list, then that would also mean all + // its prerequisites are in the list and we can just return its + // position. Unless we want it reordered. + // + iterator& pos (mi->second.position); + if (pos != end ()) + { + if (reorder) + erase (pos); + else + return pos; + } + + // Order all the prerequisites of this package and compute the position of + // its "earliest" prerequisite -- this is where it will be inserted. + // + const shared_ptr& sp (p.selected); + const shared_ptr& ap (p.available); + + bool build (*p.action == build_package::build); + + // Package build must always have the available package associated. + // + assert (!build || ap != nullptr); + + // Unless this package needs something to be before it, add it to the end + // of the list. + // + iterator i (end ()); + + // Figure out if j is before i, in which case set i to j. The goal here is + // to find the position of our "earliest" prerequisite. + // + auto update = [this, &i] (iterator j) + { + for (iterator k (j); i != j && k != end ();) + if (++k == i) + i = j; + }; + + // Similar to collect_build(), we can prune if the package is already + // configured, right? While in collect_build() we didn't need to add + // prerequisites of such a package, it doesn't mean that they actually + // never ended up in the map via another dependency path. For example, + // some can be a part of the initial selection. And in that case we must + // order things properly. + // + // Also, if the package we are ordering is not a system one and needs to + // be disfigured during the plan execution, then we must order its + // (current) dependencies that also need to be disfigured. + // + // And yet, if the package we are ordering is a repointed dependent, then + // we must order not only its unamended and new prerequisites but also its + // replaced prerequisites, which can also be disfigured. + // + bool src_conf (sp != nullptr && + sp->state == package_state::configured && + sp->substate != package_substate::system); + + auto disfigure = [] (const build_package& p) + { + return p.action && (*p.action == build_package::drop || p.reconfigure ()); + }; + + bool order_disfigured (src_conf && disfigure (p)); + + chain.push_back (cp); + + // Order the build dependencies. + // + if (build && !p.system) + { + // So here we are going to do things differently depending on whether + // the package is already configured or not. If it is and not as a + // system package, then that means we can use its prerequisites + // list. Otherwise, we use the manifest data. + // + if (src_conf && + sp->version == p.available_version () && + (p.config_vars.empty () || + !has_buildfile_clause (ap->dependencies))) + { + for (const auto& p: sp->prerequisites) + { + database& db (p.first.database ()); + const package_name& name (p.first.object_id ()); + + // The prerequisites may not necessarily be in the map. + // + // Note that for the repointed dependent we also order its new and + // replaced prerequisites here, since they all are in the selected + // package prerequisites set. + // + auto i (map_.find (db, name)); + if (i != map_.end () && i->second.package.action) + update (order (db, + name, + nullopt /* buildtime */, + chain, + fdb, + false /* reorder */)); + } + + // We just ordered them among other prerequisites. + // + order_disfigured = false; + } + else + { + // The package prerequisites builds must already be collected and + // thus the resulting dependency list is complete. + // + assert (p.dependencies && + p.dependencies->size () == ap->dependencies.size ()); + + // We are iterating in reverse so that when we iterate over the + // dependency list (also in reverse), prerequisites will be built in + // the order that is as close to the manifest as possible. + // + for (const dependency_alternatives_ex& das: + reverse_iterate (*p.dependencies)) + { + // The specific dependency alternative must already be selected, + // unless this is a toolchain build-time dependency or all the + // alternatives are disabled in which case the alternatives list + // is empty. + // + if (das.empty ()) + continue; + + assert (das.size () == 1); + + for (const dependency& d: das.front ()) + { + // Note that for the repointed dependent we only order its new and + // unamended prerequisites here. Its replaced prerequisites will + // be ordered below. + // + update (order (pdb, + d.name, + das.buildtime, + chain, + fdb, + false /* reorder */)); + } + } + } + } + + // Order the dependencies being disfigured. + // + if (order_disfigured) + { + for (const auto& p: sp->prerequisites) + { + database& db (p.first.database ()); + const package_name& name (p.first.object_id ()); + + // The prerequisites may not necessarily be in the map. + // + auto i (map_.find (db, name)); + + // Note that for the repointed dependent we also order its replaced + // and potentially new prerequisites here (see above). The latter is + // redundant (we may have already ordered them above) but harmless, + // since we do not reorder. + // + if (i != map_.end () && disfigure (i->second.package)) + update (order (db, + name, + nullopt /* buildtime */, + chain, + fdb, + false /* reorder */)); + } + } + + chain.pop_back (); + + return pos = insert (i, p); + } + + build_packages::package_map::iterator build_packages::package_map:: + find_dependency (database& db, const package_name& pn, bool buildtime) + { + iterator r (end ()); + + linked_databases ldbs (db.dependency_configs (pn, buildtime)); + + for (database& ldb: ldbs) + { + iterator i (find (ldb, pn)); + if (i != end ()) + { + if (r == end ()) + r = i; + else + fail << "building package " << pn << " in multiple " + << "configurations" << + info << r->first.db.get().config_orig << + info << ldb.config_orig << + info << "use --config-* to select package configuration"; + } + } + + return r; + } +} -- cgit v1.1