From 30a14d230ed118363650259aca953474cd3681ed Mon Sep 17 00:00:00 2001
From: Mitchell Hashimoto <mitchell.hashimoto@gmail.com>
Date: Thu, 1 Sep 2022 17:53:40 -0700
Subject: [PATCH] process ASCII events manually to avoid function call overhead

---
 src/Window.zig            | 40 ++++++++++++++++++++++++++++++++++++---
 src/terminal/Parser.zig   |  4 ++++
 src/terminal/Screen.zig   |  4 ++++
 src/terminal/Terminal.zig |  8 +++++++-
 src/terminal/main.zig     |  1 +
 5 files changed, 53 insertions(+), 4 deletions(-)

diff --git a/src/Window.zig b/src/Window.zig
index f0407f7e8..6b1674dcf 100644
--- a/src/Window.zig
+++ b/src/Window.zig
@@ -1271,9 +1271,43 @@ fn ttyRead(t: *libuv.Tty, n: isize, buf: []const u8) void {
     // Schedule a render
     win.render_timer.schedule() catch unreachable;
 
-    // Process the terminal data
-    win.terminal_stream.nextSlice(buf[0..@intCast(usize, n)]) catch |err|
-        log.err("error processing terminal data: {}", .{err});
+    // Process the terminal data. This is an extremely hot part of the
+    // terminal emulator, so we do some abstraction leakage to avoid
+    // function calls and unnecessary logic.
+    //
+    // The ground state is the only state that we can see and print/execute
+    // ASCII, so we only execute this hot path if we're already in the ground
+    // state.
+    //
+    // Empirically, this alone improved throughput of large text output by ~20%.
+    var i: usize = 0;
+    const end = @intCast(usize, n);
+    if (win.terminal_stream.parser.state == .ground) {
+        for (buf[i..end]) |c| {
+            switch (terminal.parse_table.table[c][@enumToInt(terminal.Parser.State.ground)].action) {
+                // Print, call directly.
+                .print => win.print(@intCast(u21, c)) catch |err|
+                    log.err("error processing terminal data: {}", .{err}),
+
+                // C0 execute, let our stream handle this one but otherwise
+                // continue since we're guaranteed to be back in ground.
+                .execute => win.terminal_stream.next(c) catch |err|
+                    log.err("error processing terminal data: {}", .{err}),
+
+                // Otherwise, break out and go the slow path until we're
+                // back in ground.
+                else => break,
+            }
+
+            i += 1;
+        }
+    }
+
+    if (i < end) {
+        //log.warn("SLOW={}", .{end - i});
+        win.terminal_stream.nextSlice(buf[i..end]) catch |err|
+            log.err("error processing terminal data: {}", .{err});
+    }
 }
 
 fn ttyWrite(req: *libuv.WriteReq, status: i32) void {
diff --git a/src/terminal/Parser.zig b/src/terminal/Parser.zig
index cd77007e9..032fa42d9 100644
--- a/src/terminal/Parser.zig
+++ b/src/terminal/Parser.zig
@@ -6,6 +6,7 @@ const Parser = @This();
 
 const std = @import("std");
 const builtin = @import("builtin");
+const trace = @import("tracy").trace;
 const testing = std.testing;
 const table = @import("parse_table.zig").table;
 const osc = @import("osc.zig");
@@ -212,6 +213,9 @@ pub fn init() Parser {
 /// Up to 3 actions may need to be exected -- in order -- representing
 /// the state exit, transition, and entry actions.
 pub fn next(self: *Parser, c: u8) [3]?Action {
+    const tracy = trace(@src());
+    defer tracy.end();
+
     // If we're processing UTF-8, we handle this manually.
     if (self.state == .utf8) {
         return .{ self.next_utf8(c), null, null };
diff --git a/src/terminal/Screen.zig b/src/terminal/Screen.zig
index 38b704831..169d9171c 100644
--- a/src/terminal/Screen.zig
+++ b/src/terminal/Screen.zig
@@ -22,6 +22,7 @@ const assert = std.debug.assert;
 const Allocator = std.mem.Allocator;
 
 const utf8proc = @import("utf8proc");
+const trace = @import("tracy").trace;
 const color = @import("color.zig");
 const point = @import("point.zig");
 const CircBuf = @import("circ_buf.zig").CircBuf;
@@ -432,6 +433,9 @@ pub fn rowIterator(self: *Screen, tag: RowIndexTag) RowIterator {
 /// Returns the row at the given index. This row is writable, although
 /// only the active area should probably be written to.
 pub fn getRow(self: *Screen, index: RowIndex) Row {
+    const tracy = trace(@src());
+    defer tracy.end();
+
     // Get our offset into storage
     const offset = index.toScreen(self).screen * (self.cols + 1);
 
diff --git a/src/terminal/Terminal.zig b/src/terminal/Terminal.zig
index 8b15ef055..7650d032a 100644
--- a/src/terminal/Terminal.zig
+++ b/src/terminal/Terminal.zig
@@ -464,7 +464,7 @@ pub fn print(self: *Terminal, c: u21) !void {
 
     switch (width) {
         // Single cell is very easy: just write in the cell
-        1 => _ = self.printCell(c),
+        1 => _ = @call(.{ .modifier = .always_inline }, self.printCell, .{c}),
 
         // Wide character requires a spacer. We print this by
         // using two cells: the first is flagged "wide" and has the
@@ -505,6 +505,9 @@ pub fn print(self: *Terminal, c: u21) !void {
 }
 
 fn printCell(self: *Terminal, unmapped_c: u21) *Screen.Cell {
+    // const tracy = trace(@src());
+    // defer tracy.end();
+
     const c = c: {
         // TODO: non-utf8 handling, gr
 
@@ -558,6 +561,9 @@ fn printCell(self: *Terminal, unmapped_c: u21) *Screen.Cell {
 }
 
 fn printWrap(self: *Terminal) !void {
+    const tracy = trace(@src());
+    defer tracy.end();
+
     const row = self.screen.getRow(.{ .active = self.screen.cursor.y });
     row.setWrapped(true);
 
diff --git a/src/terminal/main.zig b/src/terminal/main.zig
index 097bade52..4472116c6 100644
--- a/src/terminal/main.zig
+++ b/src/terminal/main.zig
@@ -7,6 +7,7 @@ const csi = @import("csi.zig");
 const sgr = @import("sgr.zig");
 pub const point = @import("point.zig");
 pub const color = @import("color.zig");
+pub const parse_table = @import("parse_table.zig");
 
 pub const Charset = charsets.Charset;
 pub const CharsetSlot = charsets.Slots;