feat(v5te): implement basic DTCM + ITCM

src/arm.zig

@@ -28,6 +28,8 @@ const condition_lut = [_]u16{
 };
 
 pub fn Arm32(comptime arch: Architecture) type {
+    const is_v5te = arch == .v5te;
+
     return struct {
         const Self = @This();
 
@@ -40,6 +42,10 @@ pub fn Arm32(comptime arch: Architecture) type {
 
         bank: Bank = Bank.create(),
 
+        // The following will be `void` on.v4t but exist on v5te
+        itcm: if (is_v5te) Itcm else void,
+        dtcm: if (is_v5te) Dtcm else void,
+
         const arm = switch (arch) {
             .v4t => @import("arm/v4t.zig").arm,
             .v5te => @import("arm/v5te.zig").arm,
@@ -153,16 +159,51 @@ pub fn Arm32(comptime arch: Architecture) type {
                 .bus = bus,
                 .cpsr = .{ .raw = 0x0000_001F },
                 .spsr = .{ .raw = 0x0000_0000 },
+
+                .dtcm = if (is_v5te) .{} else {},
+                .itcm = if (is_v5te) .{} else {},
             };
         }
 
         // CPU needs it's own read/write fns due to ICTM and DCTM present in v5te
         // I considered implementing Bus.cpu_read and Bus.cpu_write but ended up considering that a bit too leaky
         pub fn read(self: *Self, comptime T: type, address: u32) T {
+            const readInt = std.mem.readIntSliceLittle;
+
+            if (is_v5te) {
+                const itcm_base: u32 = self.itcm.base_address;
+                const itcm_size = self.itcm.buf.len;
+                const dtcm_base: u32 = self.dtcm.base_address;
+                const dtcm_size = self.dtcm.buf.len;
+
+                // FIXME: verify correctness + can this be faster?
+                if (itcm_base < address and address < itcm_base + itcm_size)
+                    return readInt(T, self.itcm.buf[address & 0x0000_7FFF ..][0..@sizeOf(T)]);
+
+                if (dtcm_base < address and address < dtcm_base + dtcm_size)
+                    return readInt(T, self.itcm.buf[address & 0x0000_3FFF ..][0..@sizeOf(T)]);
+            }
+
             return self.bus.read(T, address);
         }
 
         pub fn write(self: *Self, comptime T: type, address: u32, value: T) void {
+            const writeInt = std.mem.writeIntSliceLittle;
+
+            if (is_v5te) {
+                const itcm_base: u32 = self.itcm.base_address;
+                const itcm_size = self.itcm.buf.len;
+                const dtcm_base: u32 = self.dtcm.base_address;
+                const dtcm_size = self.dtcm.buf.len;
+
+                // FIXME: verify correctness + can this be faster?
+                if (itcm_base < address and address < itcm_base + itcm_size)
+                    return writeInt(T, self.itcm.buf[address & 0x0000_7FFF ..][0..@sizeOf(T)], value);
+
+                if (dtcm_base < address and address < dtcm_base + dtcm_size)
+                    return writeInt(T, self.itcm.buf[address & 0x0000_3FFF ..][0..@sizeOf(T)], value);
+            }
+
             return self.bus.write(T, address, value);
         }
 
@@ -173,6 +214,9 @@ pub fn Arm32(comptime arch: Architecture) type {
                 .bus = self.bus,
                 .cpsr = .{ .raw = 0x0000_001F },
                 .spsr = .{ .raw = 0x0000_0000 },
+
+                .dtcm = if (is_v5te) .{} else {},
+                .itcm = if (is_v5te) .{} else {},
             };
         }
 
@@ -352,6 +396,18 @@ pub fn Arm32(comptime arch: Architecture) type {
     };
 }
 
+fn Tcm(comptime count: usize, comptime default_addr: u32) type {
+    const KiB = 0x400;
+
+    return struct {
+        buf: [count * KiB]u8 = [_]u8{0x00} ** (count * KiB),
+        base_address: u32 = default_addr,
+    };
+}
+
+const Itcm = Tcm(32, 0x0000_0000);
+const Dtcm = Tcm(16, 0x0080_0000); // GBATEK says default is 0x027C_0000...
+
 pub const Mode = enum(u5) {
     User = 0b10000,
     Fiq = 0b10001,