The long production line on which products of Inform are built, one step at a time.

§1. Having seen the top management of the factory, we now reach the factory floor, which is one long production line: over 100 steps must be performed in sequence to finish making the product. We can picture each of these steps as carried out by a worker function: each does its work and passes on to the next.

We group the steps into departments, which are in order of when they work:

enum SUBDIVIDING_CSEQ from 0
enum BUILT_IN_STUFF_CSEQ
enum SEMANTIC_ANALYSIS_CSEQ
enum ASSERTIONS_PASS_1_CSEQ
enum ASSERTIONS_PASS_2_CSEQ
enum MODEL_CSEQ
enum MODEL_COMPLETE_CSEQ
enum TABLES_CSEQ
enum AUGMENT_CSEQ
enum PHRASES_CSEQ
enum INTER1_CSEQ
enum INTER2_CSEQ
enum INTER3_CSEQ
enum INTER4_CSEQ
enum INTER5_CSEQ
enum BIBLIOGRAPHIC_CSEQ
enum FINISHED_CSEQ

§2. The aim of this section is to contain as little logic as possible other then the sequence itself.

text_stream *current_sequence_bench = NULL;

int Sequence::carry_out?Usage of Sequence::carry_out:
What To Compile - §3(int debugging) {
    stopwatch_timer *sequence_timer =
        Time::start_stopwatch(inform7_timer, I"compilation to Inter");
    current_sequence_bench = Str::new();
    Divide into compilation units2.2;
    Build a rudimentary set of kinds, relations, verbs and inference subjects2.3;
    Pass three times through the major nodes2.4;
    Make the model world2.5;
    Tables and grammar2.6;
    Augment model world with low-level properties2.7;
    Phrases and rules2.8;
    Run any internal tests2.9;
    Generate inter, part 12.10
    Generate inter, part 22.11
    Generate inter, part 32.12
    Generate inter, part 42.13
    Generate inter, part 52.14
    Generate index and bibliographic file2.15;
    if (problem_count == 0) Sequence::throw_error_if_subtasks_remain();
    Task::advance_stage_to(FINISHED_CSEQ, I"Complete", -1, debugging, sequence_timer);
    Str::clear(current_sequence_bench);
    int cpu_time_used = Time::stop_stopwatch(sequence_timer);
    LOG("Compile CPU time: %d centiseconds\n", cpu_time_used);
    if (problem_count > 0) return FALSE;
    return TRUE;
}

§2.1. This macro carries out a step at what we think of as "benches" in the production line: hence the name BENCH. Continuing the analogy, there is an ongoing time and motion study: any step which takes more than 1 centisecond of CPU time is reported to the debugging log. That isn't necessarily a sign of something wrong: a few of these steps are always going to take serious computation. But we want to know which ones.

As soon as any step generates problem messages, all subsequent steps are skipped, so that each of the worker functions can assume that the part-assembled state they receive is correct so far. This greatly simplifies error recovery, prevents small errors in the source text leading to cascades of problem messages, and ensures that we report problems as quickly as possible.

define BENCH(routine) {
    if (problem_count == 0) {
        TEMPORARY_TEXT(name)
        WRITE_TO(name, "//");
        WRITE_TO(name, #routine);
        WRITE_TO(name, "//");
        for (int i=0; i<Str::len(name)-1; i++)
            if ((Str::get_at(name, i) == '_') && (Str::get_at(name, i+1) == '_')) {
                Str::put_at(name, i, ':'); Str::put_at(name, i+1, ':');
            }
        stopwatch_timer *st = Time::start_stopwatch(sequence_timer, name);
        Str::clear(current_sequence_bench);
        WRITE_TO(current_sequence_bench, "%S", name);
        DISCARD_TEXT(name)
        routine();
        int cs = Time::stop_stopwatch(st);
        if (cs > 0) LOG(".... " #routine "() took %dcs\n", cs);
    }
}

§2.2. Here, then, are the steps in the production line, presented without commentary. For what they do, see the relevant sections. Note that at the end of each stage, plugins made by compiler features are allowed to add further steps; see Task::advance_stage_to.

Before anything else can be done, we must create an empty Inter hierarchy into which we will "emit" an Inter program. No actual code will be emitted for some time yet, but identifier names and type declarations need somewhere to go. We then break the source into "compilation units" — basically, one for the main source text and one for each extension — because the Inter hierarchy will divide according to these units.

Divide into compilation units2.2 =

    Task::advance_stage_to(SUBDIVIDING_CSEQ, I"Dividing source into compilation units",
        -1, debugging, sequence_timer);
    BENCH(CompilationSettings::initialise_gcs)
    BENCH(Emit::create_emission_tree)
    BENCH(Hierarchy::establish)
    BENCH(GenericModule::compile);
    BENCH(NameResolution::make_the_tree)
    BENCH(Task::write_XML_headings_file)
    BENCH(CompilationUnits::determine)
    BENCH(Task::warn_about_deprecated_nests)

• This code is used in §2.

§2.3. Most of the conceptual infrastructure in Inform is created by Inform source text in the Basic Inform or Standard Rules extensions, but not basic kinds of value such as "number", or the verb "to mean", or the meaning relation, and so on. Those absolute basics are made here.

Build a rudimentary set of kinds, relations, verbs and inference subjects2.3 =

    Task::advance_stage_to(BUILT_IN_STUFF_CSEQ, I"Making built in infrastructure",
        -1, debugging, sequence_timer);
    BENCH(InferenceSubjects::make_built_in);
    BENCH(Task::make_built_in_kind_constructors);
    BENCH(BinaryPredicateFamilies::first_stock)
    BENCH(BootVerbs::make_built_in)
    BENCH(Instances::make_instances_from_Neptune);

• This code is used in §2.

§2.4. Pass three times through the major nodes2.4 =

    Task::advance_stage_to(SEMANTIC_ANALYSIS_CSEQ, I"Pre-pass through major nodes",
        1, debugging, sequence_timer);
    BENCH(MajorNodes::pre_pass)
    BENCH(Task::verify)
    Task::advance_stage_to(ASSERTIONS_PASS_1_CSEQ, I"First pass through major nodes",
        2, debugging, sequence_timer);
    BENCH(MajorNodes::pass_1)
    BENCH(Tables::traverse_to_stock)
    BENCH(LiteralPatterns::after_pass_1)
    BENCH(Dialogue::after_pass_1)
    Task::advance_stage_to(ASSERTIONS_PASS_2_CSEQ, I"Second pass through major nodes",
        -1, debugging, sequence_timer);
    BENCH(MajorNodes::pass_2)
    BENCH(DialogueBeats::decide_cue_topics)
    BENCH(LicenceDeclaration::check_licences)
    BENCH(BibliographicData::fill_licence_variables)

• This code is used in §2.

§2.5. Make the model world2.5 =

    Task::advance_stage_to(MODEL_CSEQ, I"Making the model world",
        -1, debugging, sequence_timer);
    BENCH(RTKindDeclarations::declare_base_kinds)
    BENCH(Translations::traverse_for_late_namings)
    BENCH(OrderingInstances::objects_in_definition_sequence)
    Task::advance_stage_to(MODEL_COMPLETE_CSEQ, I"Completing the model world",
        -1, debugging, sequence_timer);
    BENCH(World::stages_II_and_III)
    BENCH(World::stage_IV)

• This code is used in §2.

§2.6. Tables and grammar2.6 =

    Task::advance_stage_to(TABLES_CSEQ, I"Tables and grammar",
        -1, debugging, sequence_timer);
    BENCH(Measurements::validate_definitions)
    BENCH(BinaryPredicateFamilies::second_stock)
    BENCH(Tables::check_tables_for_kind_clashes)

• This code is used in §2.

§2.7. Augment model world with low-level properties2.7 =

    Task::advance_stage_to(AUGMENT_CSEQ, I"Augment model world",
        -1, debugging, sequence_timer);
    BENCH(World::stage_V)
    BENCH(MappingHints::traverse_for_map_parameters)

• This code is used in §2.

§2.8. Phrases and rules2.8 =

    Task::advance_stage_to(PHRASES_CSEQ, I"Phrases and rules",
        3, debugging, sequence_timer);
    BENCH(LiteralPatterns::define_named_phrases)
    BENCH(ImperativeDefinitions::assess_all)
    BENCH(Equations::traverse_to_stock)
    BENCH(Tables::traverse_to_stock)
    BENCH(RTRulebooks::RulebookOutcomePrintingRule)

• This code is used in §2.

§2.9. See Internal Test Cases for an explanation of the alarming-looking exit here, which only happens when special runs are made for compiler testing.

Run any internal tests2.9 =

    if ((debugging) && (problem_count == 0)) {
        int tests_run = InternalTests::run(1);
        if (tests_run > 0) exit(0);
    }

• This code is used in §2.

§2.10. This proceeds in stages.

Generate inter, part 12.10 =

    Task::advance_stage_to(INTER1_CSEQ, I"Generating inter (1)",
        4, debugging, sequence_timer);
    BENCH(RTUseOptions::compile)
    BENCH(RTCommandGrammars::compile_non_generic_constants)
    BENCH(Interventions::make_all)
    BENCH(RTKindConstructors::assign_declaration_sequence_numbers)
    BENCH(RTKindConstructors::compile)
    BENCH(RTLiteralPatterns::compile)

• This code is used in §2.

§2.11. Generate inter, part 22.11 =

    Task::advance_stage_to(INTER2_CSEQ, I"Generating inter (2)",
        -1, debugging, sequence_timer);
    BENCH(CompletionModule::compile);
    BENCH(RTProperties::compile)
    BENCH(RTKindConstructors::compile_permissions)
    BENCH(InferenceSubjects::emit_all)
    BENCH(Tables::complete)
    BENCH(RTTables::compile)
    BENCH(RTTableColumns::compile)
    BENCH(RTEquations::compile)
    BENCH(ImperativeDefinitions::compile_first_block)
    BENCH(RTDialogueBeats::compile)
    BENCH(RTDialogueLines::compile)
    BENCH(RTDialogueChoices::compile)
    BENCH(RTRules::compile)
    BENCH(RTRulebooks::compile)
    BENCH(RTRulebooks::compile_nros)
    BENCH(RTActivities::compile)
    BENCH(RTVerbs::compile_conjugations)
    BENCH(RTVerbs::compile_forms)
    BENCH(CompilationUnits::complete_metadata);

• This code is used in §2.

§2.12. Generate inter, part 32.12 =

    Task::advance_stage_to(INTER3_CSEQ, I"Generating inter (3)",
        -1, debugging, sequence_timer);
    BENCH(PhraseRequests::invoke_to_begin)
    BENCH(Closures::compile_closures)
    BENCH(Sequence::undertake_queued_tasks)
    BENCH(RTRelations::compile)
    BENCH(RTAdjectives::compile_mdef_test_functions)
    BENCH(Sequence::undertake_queued_tasks)
    BENCH(RTPhrasebook::compile_entries);
    BENCH(RTMappingHints::compile)

• This code is used in §2.

§2.13. Generate inter, part 42.13 =

    Task::advance_stage_to(INTER4_CSEQ, I"Generating inter (4)",
        -1, debugging, sequence_timer);

• This code is used in §2.

§2.14. Generate inter, part 52.14 =

    Task::advance_stage_to(INTER5_CSEQ, I"Generating inter (5)",
        -1, debugging, sequence_timer);
    BENCH(Sequence::undertake_queued_tasks)
    BENCH(RTKindIDs::compile_structures)
    BENCH(Sequence::undertake_queued_tasks)
    BENCH(Sequence::allow_no_further_queued_tasks)
    BENCH(TheHeap::compile_configuration)
    BENCH(Rules::check_response_usages)
    BENCH(LocalParking::compile_array)
    BENCH(RTBibliographicData::IFID_text)
    BENCH(Sequence::lint_inter)

• This code is used in §2.

§2.15. Generate index and bibliographic file2.15 =

    Task::advance_stage_to(BIBLIOGRAPHIC_CSEQ, I"Bibliographic work",
        -1, debugging, sequence_timer);
    BENCH(Hierarchy::log);
    if ((debugging) && (problem_count == 0)) {
        int tests_run = InternalTests::run(2);
        if (tests_run > 0) exit(0);
    }
    BENCH(Task::specify_index_requirements);
    if (Log::aspect_switched_on(INTER_DA))
        InterSkill::set_debugging();
    if (Log::aspect_switched_on(INFORM_INTER_DA))
        TextualInter::write(DL, Emit::tree(), NULL);
    linked_list *L = Dash::phrases_to_log();
    if (L) {
        LOG("Phrase or rule bodies for which textual Inter was requested by '***':\n");
        int n = 1;
        inter_package *pack;
        LOOP_OVER_LINKED_LIST(pack, inter_package, L) {
            LOG("\n%d. Package at $6:\n", n++, pack);
            TextualInter::write_package(DL, Emit::tree(), pack);
        }
    }

• This code is used in §2.

§3. We will define just one of the above steps here, because it works in a way which breaks the pattern of doing everything just once. For one thing, it's actually called twice in the above sequence.

The issue here is that each time an imperative definition is compiled to a function, that can require other resources to be compiled in turn. The code compiled into the function body can involve calls to functions derived from other imperative definitions, or even the same one reinterpreted:

To expose (X - a value):
    say "You admire [X]."

To advertise (T - text):
    expose T;
    let the price be "the price tag of [a random number between 5 and 10] pounds";
    expose the price.

Every turn:
    advertise "a valuable antique silver coffee pot".

Phrases are compiled on demand, but rules are always demanded, so the "every turn" rule here is compiled; that requires "advertise" to be compiled; which in turn requires a form of "expose X" to be compiled for X a text. But "advertise" also needs the text substitution "the price tag of [a random number between 5 and 10] pounds" to be compiled, and that in turn creates a further function compilation in order to provide a context for execution of the phrase "a random number between 5 and 10", which in turn... and so on.

§4. The only way to be sure of handling all needs here is to keep on compiling until the process exhausts itself, and this we do with a queue of tasks to perform.¹ Suppose we have this queue:

    T1 -> T2 -> T3 -> T4 -> ...

and we are working on T2. That uncovers the need for three further tasks X1, X2, X3: those are added immediately after T2 —

    T1 -> T2 -> X1 -> X2 -> X3 -> T3 -> T4 -> ...

Thus we never reach T3 until T2 has been completely exhausted, including its secondary tasks. To get a sense of how this works in practice, try:

Include task queue in the debugging log.

• ¹ Until 2021 this process is structured as a set of coroutines rather than a queue. C does not strictly speaking support coroutines, though that hasn't stopped hackers from using assembly language to manipulate return addresses on the C call stack, and/or use Duff's device. It never quite came to that here, but it was sometimes difficult to reason about. ↩

§5. A task is abstracted as being a call to a function, called the "agent", with a pointer to the relevant data.

typedef struct compilation_subtask {
    struct compilation_subtask *caused_by;
    struct compilation_subtask *next_task;
    void (*agent)(struct compilation_subtask *);
    struct general_pointer data;
    struct parse_node *current_sentence_when_queued;
    struct text_stream *description;
    CLASS_DEFINITION
} compilation_subtask;

compilation_subtask *Sequence::new_subtask?Usage of Sequence::new_subtask:
§7(void (*agent)(struct compilation_subtask *),
    general_pointer gp, text_stream *desc) {
    compilation_subtask *t = CREATE(compilation_subtask);
    t->caused_by = NULL;
    t->next_task = NULL;
    t->agent = agent;
    t->data = gp;
    t->current_sentence_when_queued = current_sentence;
    t->description = desc;
    return t;
}

• The structure compilation_subtask is private to this section.

§6. Each call to Sequence::undertake_queued_tasks works methodically through the queue until everything is done.

The queue is a linked list of compilation_subtask objects in between first_task and last_task. (The queue is empty if and only if both are NULL.) The queue only grows, and never has items removed.

A marker called last_task_undetaken shows how much progress we have made in completing the tasks queued: so, when this is equal to last_task, there is nothing to do. Another marker called current_task is set only when a task is under way, and is NULL at all other times.

compilation_subtask *first_task = NULL, *last_task = NULL, *last_task_undetaken = NULL;
compilation_subtask *current_task = NULL, *current_horizon = NULL;
int task_queue_is_closed = FALSE;

§7. The rest of Inform, if it wants to schedule a compilation task, should call one of these two functions:

void Sequence::queue(void (*agent)(struct compilation_subtask *),
    general_pointer gp, text_stream *desc) {
    compilation_subtask *t = Sequence::new_subtask(agent, gp, desc);
    Queue the task7.1;
}

void Sequence::queue_at(void (*agent)(struct compilation_subtask *),
    general_pointer gp, text_stream *desc, parse_node *at) {
    compilation_subtask *t = Sequence::new_subtask(agent, gp, desc);
    t->current_sentence_when_queued = at;
    Queue the task7.1;
}

§7.1. New entries are inserted in the queue at two write positions:

• ● after the last_task, i.e., at the back, if no task is currently going on; or
• ● after the current_horizon marker, i.e., after the current task finishes.

In the case where we are currently in the middle of what was the last task when it started, these two positions will be the same, so we sometimes need to advance last_task even when current_horizon is set.

Queue the task7.1 =

    t->caused_by = current_task;
    if (first_task == NULL) { first_task = t; last_task = t; return; }
    if (current_horizon) {
        t->next_task = current_horizon->next_task;
        current_horizon->next_task = t;
        current_horizon = t;
    } else {
        t->next_task = NULL;
        last_task->next_task = t;
    }
    if (t->next_task == NULL) last_task = t;
    if (t->caused_by) WRITE_TO(t->description, " from [%d]", t->caused_by->allocation_id);
    else WRITE_TO(t->description, " from %S", current_sequence_bench);
    if (task_queue_is_closed) {
        LOG("offending task was: ");
        Sequence::write_task(DL, t);
        internal_error("too late to schedule further compilation tasks");
    }
    if (Log::aspect_switched_on(TASK_QUEUE_DA)) {
        LOG("queued:    ");
        Sequence::write_task(DL, t);
    }

• This code is used in §7 (twice).

§8. Here the chimes of midnight sound:

void Sequence::allow_no_further_queued_tasks?Usage of Sequence::allow_no_further_queued_tasks:
§2.14(void) {
    task_queue_is_closed = TRUE;
}

§9. So here is where the work is done, and the last_task_undetaken advances:

void Sequence::undertake_queued_tasks?Usage of Sequence::undertake_queued_tasks:
§2.12, §2.14(void) {
    compilation_subtask *t;
    do {
        t = first_task;
        if (last_task_undetaken) t = last_task_undetaken->next_task;
        if (t) {
            last_task_undetaken = t;
            compilation_subtask *save_task = current_task;
            compilation_subtask *save_horizon = current_horizon;
            parse_node *save = current_sentence;
            current_task = t;
            current_horizon = t;
            current_sentence = t->current_sentence_when_queued;
            (*(t->agent))(t);
            if (Log::aspect_switched_on(TASK_QUEUE_DA)) {
                LOG("completed: ");
                Sequence::write_task(DL, t);
            }
            current_sentence = save;
            current_task = save_task;
            current_horizon = save_horizon;
            current_sentence = save;
        }
    } while (t);
}

§10. At the end of compilation, the queue ought to be empty, but just in case:

void Sequence::throw_error_if_subtasks_remain?Usage of Sequence::throw_error_if_subtasks_remain:
§2(void) {
    if (first_task) {
        compilation_subtask *t = first_task;
        if (last_task_undetaken) t = last_task_undetaken->next_task;
        if (t) {
            Sequence::write_from(DL, t);
            internal_error("there are compilation tasks never reached");
        }
    }
}

void Sequence::backtrace?Usage of Sequence::backtrace:
Using Problems - §3.1(void) {
    compilation_subtask *t = current_task;
    int d = 0;
    while (t) {
        if (d++ == 0) LOG("During compilation task: ");
        else LOG("caused by compilation task: ");
        Sequence::write_task(DL, t);
        t = t->caused_by;
    }
    if (Str::len(current_sequence_bench) > 0)
        LOG("During bench %S\n", current_sequence_bench);
}

§11. And these are used for logging:

void Sequence::write_from?Usage of Sequence::write_from:
§10(OUTPUT_STREAM, compilation_subtask *t) {
    while (t) {
        Sequence::write_task(OUT, t);
        t = t->next_task;
    }
}

void Sequence::write_task?Usage of Sequence::write_task:
§7.1, §9, §10(OUTPUT_STREAM, compilation_subtask *t) {
    if (t == NULL) WRITE("[NULL]\n");
    else WRITE("[%d] %S\n", t->allocation_id, t->description);
}

§12. The final step is to verify that the Inter we have produced is correct:

void Sequence::lint_inter?Usage of Sequence::lint_inter:
§2.14(void) {
    InterInstruction::tree_lint(Emit::tree());
}