libghostty: attacking wasm binary size (#13830)

This should release our ReleaseFast bundle from 1.1MB to ~800KB. 

The major win is disabling logging in ReleaseFast wasm builds (~200KB).

The next is removing aggressive inlining in paths that don't make sense
for performance. Verified with benchmarks on native to not affect
anything really.

The third was really dumb: `var buf: [4096]u32 = @splat(c)` in LLVM
releasefast for wasm was lowering to 4096 separate `i32.store`...
like... 30KB of code. Replacing this with a for loop reduced by 30KB and
made REP (a rare sequence) 11x faster lol.
This commit is contained in:
Mitchell Hashimoto
2026-08-14 21:02:41 -07:00
committed by GitHub
4 changed files with 53 additions and 41 deletions

View File

@@ -391,16 +391,15 @@ pub const std_options: std.Options = opts: {
var options: std.Options = .{};
if (builtin.target.cpu.arch.isWasm()) {
// Wasm builds we specifically want to optimize for space with small
// releases so we bump up to warn. Everything else acts pretty normal.
options.log_level = switch (builtin.mode) {
.Debug => .debug,
.ReleaseSmall => .warn,
else => .info,
};
// Wasm doesn't have access to stdio so we have a custom log function.
options.logFn = @import("os/wasm/log.zig").log;
// In non-debug modes, we want to ship effectively no logging
// warn and lower add ~200KB at the time of this comment.
if (builtin.mode == .Debug) {
options.log_level = .debug;
options.logFn = @import("os/wasm/log.zig").log;
} else {
options.log_level = .err;
options.logFn = @import("os/wasm/log.zig").noop;
}
} else if (terminal.options.c_abi) {
// For C ABI builds, use a custom log function that dispatches to an
// embedder-provided callback (or silently discards when none is set).

View File

@@ -4,6 +4,22 @@ const wasm = @import("../wasm.zig");
// Use the correct implementation
pub const log = Freestanding.log;
/// A log function that discards everything. Since the format string and
/// scope are comptime parameters, using this compiles out all logging
/// machinery: call sites, format strings, and the std.fmt code they
/// reference (~220KB in a ReleaseFast wasm build).
pub fn noop(
comptime level: std.log.Level,
comptime scope: @TypeOf(.EnumLiteral),
comptime format: []const u8,
args: anytype,
) void {
_ = level;
_ = scope;
_ = format;
_ = args;
}
/// Freestanding implementation calls an extern "log" function.
pub const Freestanding = struct {
// The function std.log will call.

View File

@@ -498,7 +498,14 @@ pub fn printRepeat(self: *Terminal, count_req: usize) !void {
// identical to calling print per codepoint: ineligible characters
// or terminal states (insert mode, grapheme clustering, hyperlinks,
// etc.) fall back to the per-codepoint print() path internally.
var buf: [4096]u32 = @splat(c);
//
// The buffer is filled with a runtime-bounded loop rather than
// `= @splat(c)`: a comptime-known 4096-element splat gets fully
// unrolled into ~33KB of consecutive stores (LLVM won't re-roll
// or vectorize it, see quirks_memset.zig), and it would fill the
// whole buffer even for the typical small repeat counts.
var buf: [4096]u32 = undefined;
for (buf[0..@min(remaining, buf.len)]) |*cp| cp.* = c;
while (remaining > 0) {
const n = @min(remaining, buf.len);
try self.printSlice(buf[0..n]);

View File

@@ -1268,32 +1268,17 @@ pub const PageFormatter = struct {
if (style_id == invalid_style_id) break :fast;
}
// Specialized on point tracking so that the common
// non-tracking case has zero per-cell overhead.
const consumed = if (self.point_map == null)
try self.writeCellRun(
emit,
false,
writer,
cells_subset[cell_i..],
x,
y,
style_id,
current_hyperlink_id,
&blank_cells,
)
else
try self.writeCellRun(
emit,
true,
writer,
cells_subset[cell_i..],
x,
y,
style_id,
current_hyperlink_id,
&blank_cells,
);
const consumed = try self.writeCellRun(
emit,
self.point_map != null,
writer,
cells_subset[cell_i..],
x,
y,
style_id,
current_hyperlink_id,
&blank_cells,
);
// Zero cells consumed means the first cell isn't
// eligible for the fast path; handle it below.
@@ -1579,10 +1564,15 @@ pub const PageFormatter = struct {
/// Blank cell accounting matches the slow path: accumulated blanks
/// are only materialized once a non-blank cell is found, and any
/// remainder is written back to `blank_cells`.
fn writeCellRun(
// Deliberately not inlined: this is instantiated per emit format and
// inlining every copy into formatWithStateEmit's row loop bloats the
// binary. track_points is a runtime bool for the same reason: a
// comptime bool doubles the instantiation count for one predictable
// branch per emitted cell.
noinline fn writeCellRun(
self: *const PageFormatter,
comptime emit: Format,
comptime track_points: bool,
track_points: bool,
writer: *std.Io.Writer,
cells: []const Cell,
run_x: size.CellCountInt,
@@ -1666,7 +1656,7 @@ pub const PageFormatter = struct {
// This cell produces output: materialize accumulated blanks.
if (pending > 0) {
if (comptime track_points) try self.appendBlankPoints(
if (track_points) try self.appendBlankPoints(
&self.point_map.?,
pending,
x,
@@ -1718,7 +1708,7 @@ pub const PageFormatter = struct {
}
// All of the cell's bytes map to the cell's coordinate.
if (comptime track_points) {
if (track_points) {
const map = &self.point_map.?;
map.map.appendNTimes(
map.alloc,