//! THE FILESYSTEM SCOREBOARD: what one acme-fs request costs the core, and
//! what serving one costs an editor that nobody is scripting.
//!
//!   zig build fs-bench                  -- the table
//!   zig build fs-bench -- --json        -- the same, machine-readable
//!   zig build fs-bench -- --reps 200000 -- more samples per row
//!
//! There is no FUSE here, and no thread: `acmefs.handle` IS the transaction
//! (src/acmefs.zig), so driving it directly is measuring the whole of what the
//! core does per request. That is the point of the split — a transport adds a
//! `read(2)`, a `write(2)` and a wake, and those are the kernel's numbers, not
//! ours (the mounted end-to-end figures are measured with real clients; see
//! examples/README.md).
//!
//! THREE QUESTIONS THIS ANSWERS.
//!
//!   1. Is a request cheap enough to serve thousands per frame? Each row is
//!      one `handle` call, median of `--reps`.
//!   2. Does the steady state ALLOCATE? Every row is measured through a
//!      counting allocator and the table prints the allocation count. A
//!      non-zero number in a read row is a bug: reads answer with a range of
//!      the pane's live text (`Payload.region`) or with the staging buffer,
//!      and the staging buffer is cleared, never freed.
//!   3. What does an editor with NO script attached pay? The last two rows are
//!      the same keystroke with zero listeners and with one. Zero listeners
//!      must be indistinguishable from an editor with no filesystem compiled
//!      in at all: one branch in `file_pane.setContent`.
const std = @import("std");
const pardes = @import("pardes");
const acmefs = pardes.acmefs;

pub const std_options: std.Options = .{ .log_level = .err };

const backing = std.heap.page_allocator;

/// std.time.Timer is gone in 0.16; clock_gettime is what test/perf.zig uses.
fn nowNs() u64 {
    var ts: std.c.timespec = undefined;
    _ = std.c.clock_gettime(.MONOTONIC, &ts);
    return @as(u64, @intCast(ts.sec)) *| 1_000_000_000 +| @as(u64, @intCast(ts.nsec));
}

const Row = struct {
    name: []const u8,
    ns: u64,
    allocs: usize,
    note: []const u8 = "",
};

var rows: std.ArrayList(Row) = .empty;

fn record(name: []const u8, total_ns: u64, reps: usize, counting: *std.testing.FailingAllocator, note: []const u8) void {
    rows.append(backing, .{
        .name = name,
        .ns = total_ns / @max(1, reps),
        .allocs = counting.allocations,
        .note = note,
    }) catch {};
}

/// A session with something to measure against: one big file pane, one shell,
/// and a scripted pane whose event queue has records waiting.
const Session = struct {
    core: *pardes.Pardes,
    counting: *std.testing.FailingAllocator,
    file_id: usize,
    file_serial: u32,

    fn init(counting: *std.testing.FailingAllocator, body_bytes: usize) !Session {
        const gpa = counting.allocator();
        const core = try pardes.Pardes.init(gpa, .{ .tty_only = true, .cols = 120, .rows = 40 });
        while (core.nextEffect()) |_| {}

        // A body big enough that a copy would show up in the numbers.
        const line = "the quick brown fox jumps over the lazy dog\n";
        var content: std.ArrayList(u8) = .empty;
        while (content.items.len < body_bytes) try content.appendSlice(backing, line);
        const pane = try core.hxOpenFileContent(content.items);
        content.deinit(backing);
        const id = core.active;
        while (core.nextEffect()) |_| {}
        return .{
            .core = core,
            .counting = counting,
            .file_id = id,
            .file_serial = pane.serial,
        };
    }

    fn deinit(s: *Session) void {
        s.core.deinit();
    }

    fn node(s: *const Session, file: acmefs.PaneFile) u64 {
        return acmefs.Node.of(s.file_serial, file);
    }
};

/// One row: run `req` `reps` times and report the mean cost plus how many
/// allocations the whole run made.
///
/// The per-update scratch arena is reset each rep because that is what the
/// real path does — `Pardes.update` resets it at the end of every event, and
/// `handle` called bare would otherwise let one arena grow across a hundred
/// thousand requests and count its CHUNKS as allocations. Resetting here
/// measures the request, not the harness.
fn bench(s: *Session, name: []const u8, reps: usize, req: acmefs.Req, note: []const u8) void {
    // warm the staging buffer and any lazy index the first call builds
    for (0..64) |_| {
        _ = acmefs.handle(s.core, req);
        _ = s.core.scratch.reset(.retain_capacity);
    }
    s.counting.allocations = 0;
    const start = nowNs();
    for (0..reps) |_| {
        const reply = acmefs.handle(s.core, req);
        std.mem.doNotOptimizeAway(reply.status);
        _ = s.core.scratch.reset(.retain_capacity);
    }
    record(name, nowNs() - start, reps, s.counting, note);
}

pub fn main(init: std.process.Init) !void {
    const args = try init.minimal.args.toSlice(init.arena.allocator());
    var reps: usize = 100_000;
    var json = false;
    var i: usize = 1;
    while (i < args.len) : (i += 1) {
        if (std.mem.eql(u8, args[i], "--json")) {
            json = true;
        } else if (std.mem.eql(u8, args[i], "--reps") and i + 1 < args.len) {
            i += 1;
            reps = try std.fmt.parseInt(usize, args[i], 10);
        }
    }

    // std.testing.FailingAllocator with the default options never induces a
    // failure and counts every allocation, which is the whole of what this
    // benchmark wanted from a wrapper.
    var counting: std.testing.FailingAllocator = .init(backing, .{});
    var s = try Session.init(&counting, 1 << 20); // a 1 MiB body
    defer s.deinit();

    // ---- the three shapes of request -------------------------------------
    bench(&s, "getattr body", reps, .{
        .tag = 1,
        .op = .getattr,
        .node = s.node(.body),
    }, "stat of a 1 MiB body");
    bench(&s, "lookup ctl", reps, .{
        .tag = 2,
        .op = .lookup,
        .node = acmefs.Node.of(s.file_serial, .dir),
        .data = "ctl",
    }, "name -> node");
    bench(&s, "read body 4K", reps, .{
        .tag = 3,
        .op = .read,
        .node = s.node(.body),
        .off = 4096,
        .size = 4096,
    }, "must be zero-copy");
    bench(&s, "read body 1M", reps / 10, .{
        .tag = 4,
        .op = .read,
        .node = s.node(.body),
        .off = 0,
        .size = 1 << 20,
    }, "same cost as 4K if truly zero-copy");
    bench(&s, "read ctl", reps, .{
        .tag = 5,
        .op = .read,
        .node = s.node(.ctl),
        .size = 256,
    }, "formatted into the staging buffer");
    bench(&s, "read index", reps, .{
        .tag = 6,
        .op = .read,
        .node = @intFromEnum(acmefs.TopFile.index),
        .size = 4096,
    }, "one line per pane");
    bench(&s, "readdir root", reps, .{
        .tag = 7,
        .op = .readdir,
        .node = @intFromEnum(acmefs.TopFile.root),
        .size = 4096,
    }, "staged dirents");
    bench(&s, "read event (empty)", reps, .{
        .tag = 8,
        .op = .read,
        .node = s.node(.event),
        .size = 256,
    }, "Status.again — the blocking primitive");
    // Appending GROWS the fixture, and every append is a whole-body swap plus
    // an undo snapshot (that is the core's edit model, not this filesystem's),
    // so this row is quadratic in its own rep count against a 1 MiB body.
    // Bounded on purpose: the question is what one write costs, and 200 of
    // them answer it without spending a quarter of an hour proving that
    // appending ten megabytes a kilobyte at a time is slow.
    bench(&s, "write body 1K", @min(reps, 200), .{
        .tag = 9,
        .op = .write,
        .node = s.node(.body),
        .data = "x" ** 1024,
        .size = 1024,
    }, "append: whole-body swap + undo snapshot");

    // ---- what an unscripted editor pays ----------------------------------
    // The same keystroke, twice: with nobody listening and with one listener.
    // The first number is the honest answer to "what does this feature cost a
    // session that never uses it", and the pair is what recording costs.
    //
    // A FRESH SESSION PER ROW, and this matters: typing inserts at the cursor,
    // so the line under it grows by one character per rep, and the core's
    // per-keystroke cost is dominated by walking that line's grapheme widths
    // (measured: `file_pane.graphemeDisplayWidth` is 65% of this benchmark's
    // cycles). Reusing one session made the second row type into a body the
    // first had already lengthened, and reported a 2.8x "overhead" that was
    // entirely the fixture. Small bodies for the same reason: on the megabyte
    // fixture a keystroke costs ~40 ms whatever this filesystem does.
    const key_reps = @min(reps, 2000);
    for ([_]bool{ false, true }) |scripted| {
        var keys = try Session.init(&counting, 32 * 1024);
        defer keys.deinit();
        const pane = keys.core.panes[keys.file_id].?;
        pane.mode = .insert;
        if (scripted) {
            keys.core.fs.panes[keys.file_id].readers = 1;
            keys.core.fs.listeners = 1;
        }
        counting.allocations = 0;
        const start = nowNs();
        for (0..key_reps) |_| {
            keys.core.update(.{ .key = .{ .cp = 'x', .text = "x" } });
            while (keys.core.nextEffect()) |_| {}
        }
        record(
            if (scripted) "keystroke, 1 listener" else "keystroke, no listener",
            nowNs() - start,
            key_reps,
            &counting,
            if (scripted) "diff + record" else "one branch",
        );
    }

    var out: std.Io.Writer.Allocating = .init(backing);
    defer out.deinit();
    const w = &out.writer;
    if (json) {
        try w.writeAll("{\"rows\":[");
        for (rows.items, 0..) |r, n| {
            if (n > 0) try w.writeAll(",");
            try w.print("{{\"name\":\"{s}\",\"ns\":{d},\"allocs\":{d}}}", .{ r.name, r.ns, r.allocs });
        }
        try w.writeAll("]}\n");
    } else {
        try w.print("{s:<26} {s:>9} {s:>8}  {s}\n", .{ "request", "ns/op", "allocs", "note" });
        try w.print("{s:<26} {s:>9} {s:>8}  {s}\n", .{ "-" ** 26, "-" ** 9, "-" ** 8, "-" ** 20 });
        for (rows.items) |r|
            try w.print("{s:<26} {d:>9} {d:>8}  {s}\n", .{ r.name, r.ns, r.allocs, r.note });
    }
    try std.Io.File.stdout().writeStreamingAll(init.io, out.written());
}