perf: improve equality checks

Improve equality checks during render calls by creating bespoke eql
methods or using std.mem instead of std.meta

src/Cell.zig

@@ -1,9 +1,11 @@
+const std = @import("std");
 const Image = @import("Image.zig");
 
 char: Character = .{},
 style: Style = .{},
 link: Hyperlink = .{},
 image: ?Image.Placement = null,
+default: bool = false,
 
 /// Segment is a contiguous run of text that has a constant style
 pub const Segment = struct {
@@ -59,6 +61,43 @@ pub const Style = struct {
     reverse: bool = false,
     invisible: bool = false,
     strikethrough: bool = false,
+
+    pub fn eql(a: Style, b: Style) bool {
+        const SGRBits = packed struct {
+            bold: bool,
+            dim: bool,
+            italic: bool,
+            blink: bool,
+            reverse: bool,
+            invisible: bool,
+            strikethrough: bool,
+        };
+        const a_sgr: SGRBits = .{
+            .bold = a.bold,
+            .dim = a.dim,
+            .italic = a.italic,
+            .blink = a.blink,
+            .reverse = a.reverse,
+            .invisible = a.invisible,
+            .strikethrough = a.strikethrough,
+        };
+        const b_sgr: SGRBits = .{
+            .bold = b.bold,
+            .dim = b.dim,
+            .italic = b.italic,
+            .blink = b.blink,
+            .reverse = b.reverse,
+            .invisible = b.invisible,
+            .strikethrough = b.strikethrough,
+        };
+        const a_cast: u7 = @bitCast(a_sgr);
+        const b_cast: u7 = @bitCast(b_sgr);
+        return a_cast == b_cast and
+            Color.eql(a.fg, b.fg) and
+            Color.eql(a.bg, b.bg) and
+            Color.eql(a.ul, b.ul) and
+            a.ul_style == b.ul_style;
+    }
 };
 
 pub const Color = union(enum) {
@@ -66,6 +105,19 @@ pub const Color = union(enum) {
     index: u8,
     rgb: [3]u8,
 
+    pub fn eql(a: Color, b: Color) bool {
+        if (a == .default and b == .default)
+            return true
+        else if (a == .index and b == .index)
+            return a.index == b.index
+        else if (a == .rgb and b == .rgb)
+            return a.rgb[0] == b.rgb[0] and
+                a.rgb[1] == b.rgb[1] and
+                a.rgb[2] == b.rgb[2]
+        else
+            return false;
+    }
+
     pub fn rgbFromUint(val: u24) Color {
         const r_bits = val & 0b11111111_00000000_00000000;
         const g_bits = val & 0b00000000_11111111_00000000;

src/InternalScreen.zig

@@ -16,12 +16,19 @@ pub const InternalCell = struct {
     uri_id: std.ArrayList(u8) = undefined,
     // if we got skipped because of a wide character
     skipped: bool = false,
+    default: bool = false,
 
     pub fn eql(self: InternalCell, cell: Cell) bool {
-        return std.mem.eql(u8, self.char.items, cell.char.grapheme) and
-            std.meta.eql(self.style, cell.style) and
-            std.mem.eql(u8, self.uri.items, cell.link.uri) and
-            std.mem.eql(u8, self.uri_id.items, cell.link.params);
+        // fastpath when both cells are default
+        if (self.default and cell.default) return true;
+        // this is actually faster than std.meta.eql on the individual items.
+        // Our strings are always small, usually less than 4 bytes so the simd
+        // usage in std.mem.eql has too much overhead vs looping the bytes
+        if (!std.mem.eql(u8, self.char.items, cell.char.grapheme)) return false;
+        if (!Style.eql(self.style, cell.style)) return false;
+        if (!std.mem.eql(u8, self.uri.items, cell.link.uri)) return false;
+        if (!std.mem.eql(u8, self.uri_id.items, cell.link.params)) return false;
+        return true;
     }
 };
 
@@ -94,6 +101,7 @@ pub fn writeCell(
         log.warn("couldn't write uri_id", .{});
     };
     self.buf[i].style = cell.style;
+    self.buf[i].default = cell.default;
 }
 
 pub fn readCell(self: *InternalScreen, col: usize, row: usize) ?Cell {

src/Vaxis.zig

@@ -323,7 +323,7 @@ pub fn render(self: *Vaxis) !void {
         // find out what
 
         // foreground
-        if (!std.meta.eql(cursor.fg, cell.style.fg)) {
+        if (!Cell.Color.eql(cursor.fg, cell.style.fg)) {
             const writer = tty.buffered_writer.writer();
             switch (cell.style.fg) {
                 .default => _ = try tty.write(ctlseqs.fg_reset),
@@ -340,7 +340,7 @@ pub fn render(self: *Vaxis) !void {
             }
         }
         // background
-        if (!std.meta.eql(cursor.bg, cell.style.bg)) {
+        if (!Cell.Color.eql(cursor.bg, cell.style.bg)) {
             const writer = tty.buffered_writer.writer();
             switch (cell.style.bg) {
                 .default => _ = try tty.write(ctlseqs.bg_reset),
@@ -357,7 +357,7 @@ pub fn render(self: *Vaxis) !void {
             }
         }
         // underline color
-        if (!std.meta.eql(cursor.ul, cell.style.ul)) {
+        if (!Cell.Color.eql(cursor.ul, cell.style.ul)) {
             const writer = tty.buffered_writer.writer();
             switch (cell.style.bg) {
                 .default => _ = try tty.write(ctlseqs.ul_reset),
@@ -370,7 +370,7 @@ pub fn render(self: *Vaxis) !void {
             }
         }
         // underline style
-        if (!std.meta.eql(cursor.ul_style, cell.style.ul_style)) {
+        if (cursor.ul_style != cell.style.ul_style) {
             const seq = switch (cell.style.ul_style) {
                 .off => ctlseqs.ul_off,
                 .single => ctlseqs.ul_single,
@@ -445,7 +445,7 @@ pub fn render(self: *Vaxis) !void {
         }
 
         // url
-        if (!std.meta.eql(link.uri, cell.link.uri)) {
+        if (!std.mem.eql(u8, link.uri, cell.link.uri)) {
             var ps = cell.link.params;
             if (cell.link.uri.len == 0) {
                 // Empty out the params no matter what if we don't have

src/Window.zig

@@ -196,7 +196,7 @@ pub fn writeCell(self: Window, col: usize, row: usize, cell: Cell) void {
 
 /// fills the window with the default cell
 pub fn clear(self: Window) void {
-    self.fill(.{});
+    self.fill(.{ .default = true });
 }
 
 /// returns the width of the grapheme. This depends on the terminal capabilities