`define REG_A 0
`define REG_B 1
`define REG_C 2
`define REG_D 3
`define REG_E 4
`define REG_F 5
`define REG_H 6
`define REG_L 7
`define REG_SPH 8
`define REG_SPL 9
`define REG_PCH 10
`define REG_PCL 11

`define FLAG_Z 8'b10000000
`define FLAG_N 8'b01000000
`define FLAG_H 8'b00100000
`define FLAG_C 8'b00010000

`define STATE_FETCH			2'h0
`define STATE_DECODE			2'h1
`define STATE_EXECUTE		2'h2
`define STATE_WRITEBACK		2'h3

`define INSN_LD_reg_imm8	8'b00xxx110
`define INSN_HALT				8'b01110110
`define INSN_LD_HL_reg		8'b01110xxx
`define INSN_LD_reg_HL		8'b01xxx110
`define INSN_LD_reg_reg		8'b01xxxxxx
`define INSN_LD_reg_imm16	8'b00xx0001
`define INSN_LD_SP_HL		8'b11111001
`define INSN_PUSH_reg		8'b11xx0101
`define INSN_POP_reg			8'b11xx0001
`define INSN_LDH_AC			8'b111x0010	// Either LDH A,(C) or LDH (C),A
`define INSN_LDx_AHL			8'b001xx010	// LDD/LDI A,(HL) / (HL),A
`define INSN_ALU8				8'b10xxxxxx	// 10 xxx yyy

`define INSN_reg_A		3'b111
`define INSN_reg_B		3'b000
`define INSN_reg_C		3'b001
`define INSN_reg_D		3'b010
`define INSN_reg_E		3'b011
`define INSN_reg_H		3'b100
`define INSN_reg_L		3'b101
`define INSN_reg_dHL	3'b110
`define INSN_reg16_BC	2'b00
`define INSN_reg16_DE	2'b01
`define INSN_reg16_HL	2'b10
`define INSN_reg16_SP	2'b11
`define INSN_stack_AF	2'b11
`define INSN_stack_BC	2'b00
`define INSN_stack_DE	2'b01
`define INSN_stack_HL	2'b10
`define INSN_alu_ADD		3'b000
`define INSN_alu_ADC		3'b001
`define INSN_alu_SUB		3'b010
`define INSN_alu_SBC		3'b011
`define INSN_alu_AND		3'b100
`define INSN_alu_XOR		3'b101
`define INSN_alu_OR		3'b110
`define INSN_alu_CP		3'b111		// Oh lawd, is dat some CP?

module GBZ80Core(
	input clk,
	output reg [15:0] busaddress,	/* BUS_* is latched on STATE_FETCH. */
	inout [7:0] busdata,
	output reg buswr, output reg busrd);
	
	reg [1:0] state = 0;					/* State within this bus cycle (see STATE_*). */
	reg [2:0] cycle = 0;					/* Cycle for instructions. */
	
	reg [7:0] registers[11:0];
	
	reg [15:0] address;				/* Address for the next bus operation. */
	
	reg [7:0] opcode;				/* Opcode from the current machine cycle. */
	
	reg [7:0] rdata, wdata;		/* Read data from this bus cycle, or write data for the next. */
	reg rd = 1, wr = 0, newcycle = 1;
	
	reg [7:0] tmp;					/* Generic temporary reg. */
	
	reg [7:0] buswdata;
	assign busdata = buswr ? buswdata : 8'bzzzzzzzz;
	
	initial begin
		registers[ 0] <= 0;
		registers[ 1] <= 0;
		registers[ 2] <= 0;
		registers[ 3] <= 0;
		registers[ 4] <= 0;
		registers[ 5] <= 0;
		registers[ 6] <= 0;
		registers[ 7] <= 0;
		registers[ 8] <= 0;
		registers[ 9] <= 0;
		registers[10] <= 0;
		registers[11] <= 0;
	end

	always @(posedge clk)
		case (state)
		`STATE_FETCH: begin
			if (wr)
				buswdata <= wdata;
			if (newcycle)
				busaddress <= {registers[`REG_PCH], registers[`REG_PCL]};
			else
				busaddress <= address;
			buswr <= wr;
			busrd <= rd;
			state <= `STATE_DECODE;
		end
		`STATE_DECODE: begin
			if (newcycle) begin
				opcode <= busdata;
				rdata <= busdata;
				newcycle <= 0;
				cycle <= 0;
			end else
				if (rd) rdata <= busdata;
			buswr <= 0;
			busrd <= 0;
			wr <= 0;
			rd <= 0;
			address <= 16'bxxxxxxxxxxxxxxxx;	// Make it obvious if something of type has happened.
			wdata <= 8'bxxxxxxxx;
			state <= `STATE_EXECUTE;
		end
		`STATE_EXECUTE: begin
`define EXEC_INC_PC \
	{registers[`REG_PCH], registers[`REG_PCL]} <= {registers[`REG_PCH], registers[`REG_PCL]} + 1
`define EXEC_NEXTADDR_PCINC \
	address <= {registers[`REG_PCH], registers[`REG_PCL]} + 1
`define EXEC_NEWCYCLE \
	newcycle <= 1; rd <= 1; wr <= 0
			casex (opcode)
			`INSN_LD_reg_imm8: begin
				case (cycle)
				0:	begin
						`EXEC_INC_PC;
						`EXEC_NEXTADDR_PCINC;
						rd <= 1;
					end
				1: begin
						`EXEC_INC_PC;
						if (opcode[5:3] == `INSN_reg_dHL) begin
							address <= {registers[`REG_H], registers[`REG_L]};
							wdata <= rdata;
							rd <= 0;
							wr <= 1;
						end else begin
							`EXEC_NEWCYCLE;
						end
					end
				2: begin
						`EXEC_NEWCYCLE;
					end
				endcase
			end
			`INSN_HALT: begin
				`EXEC_NEWCYCLE;
				/* XXX Interrupts needed for HALT. */
			end
			`INSN_LD_HL_reg: begin
				case (cycle)
				0:	begin
						case (opcode[2:0])
						`INSN_reg_A:	begin wdata <= registers[`REG_A]; end
						`INSN_reg_B:	begin wdata <= registers[`REG_B]; end
						`INSN_reg_C:	begin wdata <= registers[`REG_C]; end
						`INSN_reg_D:	begin wdata <= registers[`REG_D]; end
						`INSN_reg_E:	begin wdata <= registers[`REG_E]; end
						`INSN_reg_H:	begin wdata <= registers[`REG_H]; end
						`INSN_reg_L:	begin wdata <= registers[`REG_L]; end
						endcase
						address <= {registers[`REG_H], registers[`REG_L]};
						wr <= 1; rd <= 0;
					end
				1:	begin
						`EXEC_INC_PC;
						`EXEC_NEWCYCLE;
					end
				endcase
			end
			`INSN_LD_reg_HL: begin
				case(cycle)
				0: begin
						address <= {registers[`REG_H], registers[`REG_L]};
						rd <= 1;
					end
				1: begin
						tmp <= rdata;
						`EXEC_INC_PC;
						`EXEC_NEWCYCLE;
					end
				endcase
			end
			`INSN_LD_reg_reg: begin
				`EXEC_INC_PC;
				`EXEC_NEWCYCLE;
				case (opcode[2:0])
				`INSN_reg_A:	begin tmp <= registers[`REG_A]; end
				`INSN_reg_B:	begin tmp <= registers[`REG_B]; end
				`INSN_reg_C:	begin tmp <= registers[`REG_C]; end
				`INSN_reg_D:	begin tmp <= registers[`REG_D]; end
				`INSN_reg_E:	begin tmp <= registers[`REG_E]; end
				`INSN_reg_H:	begin tmp <= registers[`REG_H]; end
				`INSN_reg_L:	begin tmp <= registers[`REG_L]; end
				endcase
			end
			`INSN_LD_reg_imm16: begin
				`EXEC_INC_PC;
				case (cycle)
				0:	begin
						`EXEC_NEXTADDR_PCINC;
						rd <= 1;
					end
				1:	begin
						`EXEC_NEXTADDR_PCINC;
						rd <= 1;
					end
				2: begin `EXEC_NEWCYCLE; end
				endcase
			end
			`INSN_LD_SP_HL: begin
				case (cycle)
				0:	begin
						tmp <= registers[`REG_H];
					end
				1:	begin
						`EXEC_NEWCYCLE;
						`EXEC_INC_PC;
						tmp <= registers[`REG_L];
					end
				endcase
			end
			`INSN_PUSH_reg: begin	/* PUSH is 16 cycles! */
				case (cycle)
				0: begin
						wr <= 1;
						address <= {registers[`REG_SPH],registers[`REG_SPL]}-1;
						case (opcode[5:4])
						`INSN_stack_AF:	wdata <= registers[`REG_A];
						`INSN_stack_BC:	wdata <= registers[`REG_B];
						`INSN_stack_DE:	wdata <= registers[`REG_D];
						`INSN_stack_HL:	wdata <= registers[`REG_H];
						endcase
					end
				1: begin
						wr <= 1;
						address <= {registers[`REG_SPH],registers[`REG_SPL]}-1;
						case (opcode[5:4])
						`INSN_stack_AF:	wdata <= registers[`REG_F];
						`INSN_stack_BC:	wdata <= registers[`REG_C];
						`INSN_stack_DE:	wdata <= registers[`REG_E];
						`INSN_stack_HL:	wdata <= registers[`REG_L];
						endcase
					end
				2:	begin /* TWIDDLE OUR FUCKING THUMBS! */ end
				3: begin
						`EXEC_NEWCYCLE;
						`EXEC_INC_PC;
					end
				endcase
			end
			`INSN_POP_reg: begin	/* POP is 12 cycles! */
				case (cycle)
				0: begin
						rd <= 1;
						address <= {registers[`REG_SPH],registers[`REG_SPL]};
					end
				1: begin
						rd <= 1;
						address <= {registers[`REG_SPH],registers[`REG_SPL]};
					end
				2: begin
						`EXEC_NEWCYCLE;
						`EXEC_INC_PC;
					end
				endcase
			end
			`INSN_LDH_AC: begin
				case (cycle)
				0:	begin
						address <= {8'hFF,registers[`REG_C]};
						if (opcode[4]) begin	// LD A,(C)
							rd <= 1;
						end else begin
							wr <= 1;
							wdata <= registers[`REG_A];
						end
					end
				1: begin
						`EXEC_NEWCYCLE;
						`EXEC_INC_PC;
					end
				endcase
			end
			`INSN_LDx_AHL: begin
				case (cycle)
				0: begin
						address <= {registers[`REG_H],registers[`REG_L]};
						if (opcode[3]) begin	// LDx A, (HL)
							rd <= 1;
						end else begin
							wr <= 1;
							wdata <= registers[`REG_A];
						end
					end
				1:	begin
						`EXEC_NEWCYCLE;
						`EXEC_INC_PC;
					end
				endcase
			end
			`INSN_ALU8: begin
				if ((opcode[2:0] == `INSN_reg_dHL) && (cycle == 0)) begin
					// fffffffff fuck your shit, read from (HL) :(
					rd <= 1;
					address <= {registers[`REG_H], registers[`REG_L]};
				end else begin
					`EXEC_NEWCYCLE;
					`EXEC_INC_PC;
					case (opcode[2:0])
					`INSN_reg_A:	begin tmp <= registers[`REG_A]; end
					`INSN_reg_B:	begin tmp <= registers[`REG_B]; end
					`INSN_reg_C:	begin tmp <= registers[`REG_C]; end
					`INSN_reg_D:	begin tmp <= registers[`REG_D]; end
					`INSN_reg_E:	begin tmp <= registers[`REG_E]; end
					`INSN_reg_H:	begin tmp <= registers[`REG_H]; end
					`INSN_reg_L:	begin tmp <= registers[`REG_L]; end
					`INSN_reg_dHL:	begin tmp <= rdata; end
					endcase
				end
			end
			default:
				$stop;
			endcase
			state <= `STATE_WRITEBACK;
		end
		`STATE_WRITEBACK: begin
			casex (opcode)
			`INSN_LD_reg_imm8:
				case (cycle)
				0: cycle <= 1;
				1: case (opcode[5:3])
					`INSN_reg_A:	begin registers[`REG_A] <= rdata; cycle <= 0; end
					`INSN_reg_B:	begin registers[`REG_B] <= rdata; cycle <= 0; end
					`INSN_reg_C:	begin registers[`REG_C] <= rdata; cycle <= 0; end
					`INSN_reg_D:	begin registers[`REG_D] <= rdata; cycle <= 0; end
					`INSN_reg_E:	begin registers[`REG_E] <= rdata; cycle <= 0; end
					`INSN_reg_H:	begin registers[`REG_H] <= rdata; cycle <= 0; end
					`INSN_reg_L:	begin registers[`REG_L] <= rdata; cycle <= 0; end
					`INSN_reg_dHL:	cycle <= 2;
					endcase
				2: cycle <= 0;
				endcase
			`INSN_HALT: begin
				/* Nothing needs happen here. */
				/* XXX Interrupts needed for HALT. */
			end
			`INSN_LD_HL_reg: begin
				case (cycle)
				0: cycle <= 1;
				1: cycle <= 0;
				endcase
			end
			`INSN_LD_reg_HL: begin
				case (cycle)
				0:	cycle <= 1;
				1:	begin
						case (opcode[5:3])
						`INSN_reg_A:	begin registers[`REG_A] <= tmp; end
						`INSN_reg_B:	begin registers[`REG_B] <= tmp; end
						`INSN_reg_C:	begin registers[`REG_C] <= tmp; end
						`INSN_reg_D:	begin registers[`REG_D] <= tmp; end
						`INSN_reg_E:	begin registers[`REG_E] <= tmp; end
						`INSN_reg_H:	begin registers[`REG_H] <= tmp; end
						`INSN_reg_L:	begin registers[`REG_L] <= tmp; end
						endcase
						cycle <= 0;
					end
				endcase
			end
			`INSN_LD_reg_reg: begin
				case (opcode[5:3])
				`INSN_reg_A:	begin registers[`REG_A] <= tmp; end
				`INSN_reg_B:	begin registers[`REG_B] <= tmp; end
				`INSN_reg_C:	begin registers[`REG_C] <= tmp; end
				`INSN_reg_D:	begin registers[`REG_D] <= tmp; end
				`INSN_reg_E:	begin registers[`REG_E] <= tmp; end
				`INSN_reg_H:	begin registers[`REG_H] <= tmp; end
				`INSN_reg_L:	begin registers[`REG_L] <= tmp; end
				endcase
			end
			`INSN_LD_reg_imm16: begin
				case (cycle)
				0:	cycle <= 1;
				1:	begin
						case (opcode[5:4])
						`INSN_reg16_BC:	registers[`REG_C] <= rdata;
						`INSN_reg16_DE:	registers[`REG_E] <= rdata;
						`INSN_reg16_HL:	registers[`REG_L] <= rdata;
						`INSN_reg16_SP:	registers[`REG_SPL] <= rdata;
						endcase
						cycle <= 2;
					end
				2: begin
						case (opcode[5:4])
						`INSN_reg16_BC:	registers[`REG_B] <= rdata;
						`INSN_reg16_DE:	registers[`REG_D] <= rdata;
						`INSN_reg16_HL:	registers[`REG_H] <= rdata;
						`INSN_reg16_SP:	registers[`REG_SPH] <= rdata;
						endcase
						cycle <= 0;
					end
				endcase
			end
			`INSN_LD_SP_HL: begin
				case (cycle)
				0: begin
						cycle <= 1;
						registers[`REG_SPH] <= tmp;
					end
				1: begin
						cycle <= 0;
						registers[`REG_SPL] <= tmp;
					end
				endcase
			end
			`INSN_PUSH_reg: begin	/* PUSH is 16 cycles! */
				case (cycle)
				0: begin
						{registers[`REG_SPH],registers[`REG_SPL]} <=
							{registers[`REG_SPH],registers[`REG_SPL]} - 1;
						cycle <= 1;
					end
				1:	begin
						{registers[`REG_SPH],registers[`REG_SPL]} <=
							{registers[`REG_SPH],registers[`REG_SPL]} - 1;
						cycle <= 2;
					end
				2:	cycle <= 3;
				3:	cycle <= 0;
				endcase
			end
			`INSN_POP_reg: begin	/* POP is 12 cycles! */
				case (cycle)
				0:	begin
						cycle <= 1;
						{registers[`REG_SPH],registers[`REG_SPL]} <=
							{registers[`REG_SPH],registers[`REG_SPL]} + 1;
					end
				1:	begin
						case (opcode[5:4])
						`INSN_stack_AF:	registers[`REG_F] <= rdata;
						`INSN_stack_BC:	registers[`REG_C] <= rdata;
						`INSN_stack_DE:	registers[`REG_E] <= rdata;
						`INSN_stack_HL:	registers[`REG_L] <= rdata;
						endcase
						{registers[`REG_SPH],registers[`REG_SPL]} <=
							{registers[`REG_SPH],registers[`REG_SPL]} + 1;
						cycle <= 2;
					end
				2:	begin
						case (opcode[5:4])
						`INSN_stack_AF:	registers[`REG_A] <= rdata;
						`INSN_stack_BC:	registers[`REG_B] <= rdata;
						`INSN_stack_DE:	registers[`REG_D] <= rdata;
						`INSN_stack_HL:	registers[`REG_H] <= rdata;
						endcase
						cycle <= 0;
					end
				endcase
			end
			`INSN_LDH_AC: begin
				case (cycle)
				0:	cycle <= 1;
				1: begin
						cycle <= 0;
						if (opcode[4])
							registers[`REG_A] <= rdata;
					end
				endcase
			end
			`INSN_LDx_AHL: begin
				case (cycle)
				0:	cycle <= 1;
				1:	begin
						cycle <= 0;
						if (opcode[3])
							registers[`REG_A] <= rdata;
						{registers[`REG_H],registers[`REG_L]} <=
							opcode[4] ? // if set, LDD, else LDI
							({registers[`REG_H],registers[`REG_L]} - 1) :
							({registers[`REG_H],registers[`REG_L]} + 1);
					end
				endcase
			end
			`INSN_ALU8: begin
				if ((opcode[2:0] == `INSN_reg_dHL) && (cycle == 0)) begin
					/* Sit on our asses. */
					cycle <= 1;
				end else begin		/* Actually do the computation! */
					case (opcode[5:3])
					`INSN_alu_ADD: begin
						registers[`REG_A] <=
							registers[`REG_A] + tmp;
						registers[`REG_F] <=
							{ /* Z */ ((registers[`REG_A] + tmp) == 0) ? 1'b1 : 1'b0,
							  /* N */ 0,
							  /* H */ (({1'b0,registers[`REG_A][3:0]} + {1'b0,tmp[3:0]}) >> 4 == 1) ? 1'b1 : 1'b0,
							  /* C */ (({1'b0,registers[`REG_A]} + {1'b0,tmp}) >> 8 == 1) ? 1'b1 : 1'b0,
							  registers[`REG_F][3:0]
							};
					end
					`INSN_alu_ADC: begin
						registers[`REG_A] <=
							registers[`REG_A] + tmp + {7'b0,registers[`REG_F][4]};
						registers[`REG_F] <=
							{ /* Z */ ((registers[`REG_A] + tmp + {7'b0,registers[`REG_F][4]}) == 0) ? 1'b1 : 1'b0,
							  /* N */ 0,
							  /* H */ (({1'b0,registers[`REG_A][3:0]} + {1'b0,tmp[3:0]} + {4'b0,registers[`REG_F][4]}) >> 4 == 1) ? 1'b1 : 1'b0,
							  /* C */ (({1'b0,registers[`REG_A]} + {1'b0,tmp} + {8'b0,registers[`REG_F][4]}) >> 8 == 1) ? 1'b1 : 1'b0,
							  registers[`REG_F][3:0]
							};
					end
					`INSN_alu_AND: begin
						registers[`REG_A] <=
							registers[`REG_A] & tmp;
						registers[`REG_F] <=
							{ /* Z */ ((registers[`REG_A] & tmp) == 0) ? 1'b1 : 1'b0,
							  0,1,0,
							  registers[`REG_F][3:0]
							};
					end
					`INSN_alu_OR: begin
						registers[`REG_A] <=
							registers[`REG_A] | tmp;
						registers[`REG_F] <=
							{ /* Z */ ((registers[`REG_A] | tmp) == 0) ? 1'b1 : 1'b0,
							  0,0,0,
							  registers[`REG_F][3:0]
							};
					end
					`INSN_alu_XOR: begin
						registers[`REG_A] <=
							registers[`REG_A] ^ tmp;
						registers[`REG_F] <=
							{ /* Z */ ((registers[`REG_A] ^ tmp) == 0) ? 1'b1 : 1'b0,
							  0,0,0,
							  registers[`REG_F][3:0]
							};
					end
					default:
						$stop;
					endcase
				end
			end
			endcase
			state <= `STATE_FETCH;
		end
		endcase
endmodule

`timescale 1ns / 1ps
module TestBench();
	reg clk = 0;
	wire [15:0] addr;
	wire [7:0] data;
	wire wr, rd;
	reg [7:0] rom [2047:0];
	
	initial $readmemh("rom.hex", rom);
	always #10 clk <= ~clk;
	GBZ80Core core(
		.clk(clk),
		.busaddress(addr),
		.busdata(data),
		.buswr(wr),
		.busrd(rd));
	assign data = rd ? rom[addr] : 8'bzzzzzzzz;
endmodule