fix(input): make use of atomic rmw intrinsics when handling input

src/core/bus/io.zig

@@ -96,7 +96,7 @@ pub fn read(bus: *const Bus, comptime T: type, address: u32) ?T {
             0x0400_0128 => util.io.read.todo(log, "Read {} from SIOCNT", .{T}),
 
             // Keypad Input
-            0x0400_0130 => bus.io.keyinput.load(.Monotonic).raw,
+            0x0400_0130 => bus.io.keyinput.load(.Monotonic),
 
             // Serial Communication 2
             0x0400_0134 => util.io.read.todo(log, "Read {} from RCNT", .{T}),
@@ -366,7 +366,7 @@ const InterruptEnable = extern union {
 
 /// Read Only
 /// 0 = Pressed, 1 = Released
-const KeyInput = extern union {
+pub const KeyInput = extern union {
     a: Bit(u16, 0),
     b: Bit(u16, 1),
     select: Bit(u16, 2),
@@ -390,18 +390,19 @@ const AtomicKeyInput = struct {
         return .{ .inner = value };
     }
 
-    pub inline fn load(self: *const Self, comptime ordering: Ordering) KeyInput {
-        return .{ .raw = switch (ordering) {
+    pub inline fn load(self: *const Self, comptime ordering: Ordering) u16 {
+        return switch (ordering) {
             .AcqRel, .Release => @compileError("not supported for atomic loads"),
             else => @atomicLoad(u16, &self.inner.raw, ordering),
-        } };
+        };
     }
 
-    pub inline fn store(self: *Self, value: u16, comptime ordering: Ordering) void {
-        switch (ordering) {
-            .AcqRel, .Acquire => @compileError("not supported for atomic stores"),
-            else => @atomicStore(u16, &self.inner.raw, value, ordering),
+    pub inline fn fetchOr(self: *Self, value: u16, comptime ordering: Ordering) void {
+        _ = @atomicRmw(u16, &self.inner.raw, .Or, value, ordering);
     }
+
+    pub inline fn fetchAnd(self: *Self, value: u16, comptime ordering: Ordering) void {
+        _ = @atomicRmw(u16, &self.inner.raw, .And, value, ordering);
     }
 };
 

src/core/emu.zig

@@ -140,6 +140,10 @@ fn audioSync(audio_sync: bool, stream: *SDL.SDL_AudioStream, is_buffer_full: *bo
     // If Busy is false, there's no need to sync here
     if (!still_full) return;
 
+    // TODO: Refactor!!!!
+    // while (SDL.SDL_AudioStreamAvailable(stream) > sample_size * max_buf_size >> 1)
+    //     std.atomic.spinLoopHint();
+
     while (true) {
         still_full = SDL.SDL_AudioStreamAvailable(stream) > sample_size * max_buf_size >> 1;
         if (!audio_sync or !still_full) break;
@@ -184,7 +188,8 @@ fn sleep(timer: *Timer, wake_time: u64) ?u64 {
 }
 
 fn spinLoop(timer: *Timer, wake_time: u64) void {
-    while (true) if (timer.read() > wake_time) break;
+    while (timer.read() < wake_time)
+        std.atomic.spinLoopHint();
 }
 
 pub const EmuThing = struct {

src/imgui.zig

@@ -12,6 +12,7 @@ const emu = @import("core/emu.zig");
 const Gui = @import("platform.zig").Gui;
 const Arm7tdmi = @import("core/cpu.zig").Arm7tdmi;
 const RingBuffer = @import("zba-util").RingBuffer;
+const Dimensions = @import("platform.zig").Dimensions;
 
 const Allocator = std.mem.Allocator;
 const GLuint = gl.GLuint;
@@ -66,8 +67,10 @@ pub const State = struct {
     }
 };
 
-pub fn draw(state: *State, tex_id: GLuint, cpu: *Arm7tdmi) void {
-    const win_scale = config.config().host.win_scale;
+pub fn draw(state: *State, win_dim: Dimensions, tex_id: GLuint, cpu: *Arm7tdmi) bool {
+    const scn_scale = config.config().host.win_scale;
+
+    zgui.backend.newFrame(@intToFloat(f32, win_dim.width), @intToFloat(f32, win_dim.height));
 
     {
         _ = zgui.beginMainMenuBar();
@@ -142,8 +145,8 @@ pub fn draw(state: *State, tex_id: GLuint, cpu: *Arm7tdmi) void {
     }
 
     {
-        const w = @intToFloat(f32, gba_width * win_scale);
-        const h = @intToFloat(f32, gba_height * win_scale);
+        const w = @intToFloat(f32, gba_width * scn_scale);
+        const h = @intToFloat(f32, gba_height * scn_scale);
 
         const window_title = std.mem.sliceTo(&state.title, 0);
         _ = zgui.begin(window_title, .{ .flags = .{ .no_resize = true, .always_auto_resize = true } });
@@ -318,6 +321,8 @@ pub fn draw(state: *State, tex_id: GLuint, cpu: *Arm7tdmi) void {
     {
         zgui.showDemoWindow(null);
     }
+
+    return true; // request redraw
 }
 
 const widgets = struct {

src/platform.zig

@@ -12,6 +12,7 @@ const Arm7tdmi = @import("core/cpu.zig").Arm7tdmi;
 const Scheduler = @import("core/scheduler.zig").Scheduler;
 const FpsTracker = @import("util.zig").FpsTracker;
 const TwoWayChannel = @import("zba-util").TwoWayChannel;
+const KeyInput = @import("core/bus/io.zig").KeyInput;
 
 const gba_width = @import("core/ppu.zig").width;
 const gba_height = @import("core/ppu.zig").height;
@@ -21,7 +22,7 @@ const GLsizei = gl.GLsizei;
 const SDL_GLContext = *anyopaque;
 const Allocator = std.mem.Allocator;
 
-const Dimensions = struct { width: u32, height: u32 };
+pub const Dimensions = struct { width: u32, height: u32 };
 const default_dim: Dimensions = .{ .width = 1280, .height = 720 };
 
 pub const sample_rate = 1 << 15;
@@ -134,28 +135,9 @@ pub const Gui = struct {
                 switch (event.type) {
                     SDL.SDL_QUIT => break :emu_loop,
                     SDL.SDL_KEYDOWN => {
+                        // TODO: Make use of compare_and_xor?
                         const key_code = event.key.keysym.sym;
-                        var keyinput = cpu.bus.io.keyinput.load(.Monotonic);
-
-                        switch (key_code) {
-                            SDL.SDLK_UP => keyinput.up.unset(),
-                            SDL.SDLK_DOWN => keyinput.down.unset(),
-                            SDL.SDLK_LEFT => keyinput.left.unset(),
-                            SDL.SDLK_RIGHT => keyinput.right.unset(),
-                            SDL.SDLK_x => keyinput.a.unset(),
-                            SDL.SDLK_z => keyinput.b.unset(),
-                            SDL.SDLK_a => keyinput.shoulder_l.unset(),
-                            SDL.SDLK_s => keyinput.shoulder_r.unset(),
-                            SDL.SDLK_RETURN => keyinput.start.unset(),
-                            SDL.SDLK_RSHIFT => keyinput.select.unset(),
-                            else => {},
-                        }
-
-                        cpu.bus.io.keyinput.store(keyinput.raw, .Monotonic);
-                    },
-                    SDL.SDL_KEYUP => {
-                        const key_code = event.key.keysym.sym;
-                        var keyinput = cpu.bus.io.keyinput.load(.Monotonic);
+                        var keyinput: KeyInput = .{ .raw = 0x0000 };
 
                         switch (key_code) {
                             SDL.SDLK_UP => keyinput.up.set(),
@@ -171,7 +153,28 @@ pub const Gui = struct {
                             else => {},
                         }
 
-                        cpu.bus.io.keyinput.store(keyinput.raw, .Monotonic);
+                        cpu.bus.io.keyinput.fetchAnd(~keyinput.raw, .Monotonic);
+                    },
+                    SDL.SDL_KEYUP => {
+                        // TODO: Make use of compare_and_xor?
+                        const key_code = event.key.keysym.sym;
+                        var keyinput: KeyInput = .{ .raw = 0x0000 };
+
+                        switch (key_code) {
+                            SDL.SDLK_UP => keyinput.up.set(),
+                            SDL.SDLK_DOWN => keyinput.down.set(),
+                            SDL.SDLK_LEFT => keyinput.left.set(),
+                            SDL.SDLK_RIGHT => keyinput.right.set(),
+                            SDL.SDLK_x => keyinput.a.set(),
+                            SDL.SDLK_z => keyinput.b.set(),
+                            SDL.SDLK_a => keyinput.shoulder_l.set(),
+                            SDL.SDLK_s => keyinput.shoulder_r.set(),
+                            SDL.SDLK_RETURN => keyinput.start.set(),
+                            SDL.SDLK_RSHIFT => keyinput.select.set(),
+                            else => {},
+                        }
+
+                        cpu.bus.io.keyinput.fetchOr(keyinput.raw, .Monotonic);
                     },
                     SDL.SDL_WINDOWEVENT => {
                         if (event.window.event == SDL.SDL_WINDOWEVENT_RESIZED) {
@@ -185,7 +188,7 @@ pub const Gui = struct {
                 }
             }
 
-            zgui.backend.newFrame(@intToFloat(f32, win_dim.width), @intToFloat(f32, win_dim.height));
+            var zgui_redraw: bool = false;
 
             switch (self.state.emulation) {
                 .Transition => |inner| switch (inner) {
@@ -204,19 +207,28 @@ pub const Gui = struct {
                         self.state.emulation = .Inactive;
                     },
                 },
-                .Active => {
+                .Active => skip_draw: {
                     const is_std = mode == .Standard;
 
                     if (is_std) channel.emu.push(.Pause);
                     defer if (is_std) channel.emu.push(.Resume);
 
                     switch (mode) {
-                        .Standard => {
-                            // TODO: add timeout
-                            while (true) switch (channel.gui.pop() orelse continue) {
-                                .Paused => break,
-                                .Quit => unreachable, // only signaled in debug mode
-                            };
+                        .Standard => blk: {
+                            const limit = 15; // TODO: What should this be?
+
+                            // TODO: learn more about std.atomic.spinLoopHint();
+                            for (0..limit) |_| {
+                                const message = channel.gui.pop() orelse continue;
+
+                                switch (message) {
+                                    .Paused => break :blk,
+                                    .Quit => unreachable,
+                                }
+                            }
+
+                            log.info("timed out waiting for emu thread to pause (limit: {})", .{limit});
+                            break :skip_draw;
                         },
                         .Debug => blk: {
                             switch (channel.gui.pop() orelse break :blk) {
@@ -239,11 +251,12 @@ pub const Gui = struct {
                         opengl_impl.drawScreenTexture(emu_tex, prog_id, objects, buf);
                     }
 
-                    imgui.draw(&self.state, out_tex, cpu);
+                    zgui_redraw = imgui.draw(&self.state, win_dim, out_tex, cpu);
                 },
-                .Inactive => imgui.draw(&self.state, out_tex, cpu),
+                .Inactive => zgui_redraw = imgui.draw(&self.state, win_dim, out_tex, cpu),
             }
 
+            if (zgui_redraw) {
                 // Background Colour
                 const size = zgui.io.getDisplaySize();
                 gl.viewport(0, 0, @floatToInt(GLsizei, size[0]), @floatToInt(GLsizei, size[1]));
@@ -251,6 +264,7 @@ pub const Gui = struct {
                 gl.clear(gl.COLOR_BUFFER_BIT);
 
                 zgui.backend.draw();
+            }
 
             SDL.SDL_GL_SwapWindow(self.window);
         }