X-Git-Url: http://git.joshuawise.com/firearm.git/blobdiff_plain/9516520942af5c7dd24b0323b164c4ea2bd1dc01..ad18b18ac07db57954dc102383993fba0b6b2af4:/Decode.v?ds=inline

diff --git a/Decode.v b/Decode.v
index 85aefbf..2c22746 100644
--- a/Decode.v
+++ b/Decode.v
@@ -1,211 +1,349 @@
+`include "ARM_Constants.v"
+
 module Decode(
 	input clk,
-	input [31:0] ansn,
+	input stall,
+	input [31:0] insn,
 	input [31:0] inpc,
-	input [31:0] cps_in,
+	input [31:0] incpsr,
+	input [31:0] inspsr,
 	output reg [31:0] op0,
 	output reg [31:0] op1,
 	output reg [31:0] op2,
-	output reg [31:0] cps_out,
-
-	output [3:0] regsel0,
-	output [3:0] regsel1,
-	output [3:0] regsel2,
-	input [31:0] iregs0,
-	input [31:0] iregs1,
-	input [31:0] iregs2
+	output reg carry,
+	output reg [31:0] outcpsr,
+	output reg [31:0] outspsr,
+
+	output reg [3:0] read_0,
+	output reg [3:0] read_1,
+	output reg [3:0] read_2,
+	input [31:0] rdata_0,
+	input [31:0] rdata_1,
+	input [31:0] rdata_2
 	);
 
-	wire [31:0] regs0, regs1, regs2, rpc;
-	wire [31:0] op1_res, new_cps;
+	wire [31:0] regs0, regs1, regs2;
+	reg [31:0] rpc;
+	reg [31:0] op0_out, op1_out, op2_out;
+	reg carry_out;
 
 	/* shifter stuff */
 	wire [31:0] shift_oper;
 	wire [31:0] shift_res;
 	wire shift_cflag_out;
+	wire [31:0] rotate_res;
+
+	assign regs0 = (read_0 == 4'b1111) ? rpc : rdata_0;
+	assign regs1 = (read_1 == 4'b1111) ? rpc : rdata_1;
+	assign regs2 = rdata_2; /* use regs2 for things that cannot be r15 */
+
+	IREALLYHATEARMSHIFT shift(.insn(insn),
+	                          .operand(regs1),
+	                          .reg_amt(regs2),
+	                          .cflag_in(incpsr[`CPSR_C]),
+	                          .res(shift_res),
+	                          .cflag_out(shift_cflag_out));
+
+	SuckLessRotator whirr(.oper({24'b0, insn[7:0]}),
+	                      .amt(insn[11:8]),
+	                      .res(rotate_res));
+
+	always @(*)
+		casez (insn)
+		`DECODE_ALU_MULT,		/* Multiply -- must come before ALU, because it pattern matches a specific case of ALU */
+//		`DECODE_ALU_MUL_LONG,		/* Multiply long */
+		`DECODE_ALU_MRS,		/* MRS (Transfer PSR to register) */
+		`DECODE_ALU_MSR,		/* MSR (Transfer register to PSR) */
+		`DECODE_ALU_MSR_FLAGS,		/* MSR (Transfer register or immediate to PSR, flag bits only) */
+		`DECODE_ALU_SWP,		/* Atomic swap */
+		`DECODE_ALU_BX,			/* Branch and exchange */
+		`DECODE_ALU_HDATA_REG,		/* Halfword transfer - register offset */
+		`DECODE_ALU_HDATA_IMM,		/* Halfword transfer - register offset */
+		`DECODE_LDRSTR_UNDEFINED,	/* Undefined. I hate ARM */
+		`DECODE_LDRSTR,			/* Single data transfer */
+		`DECODE_LDMSTM,			/* Block data transfer */
+		`DECODE_BRANCH,			/* Branch */
+		`DECODE_LDCSTC,			/* Coprocessor data transfer */
+		`DECODE_CDP,			/* Coprocessor data op */
+		`DECODE_SWI:			/* SWI */
+			rpc = inpc + 8;
+		`DECODE_MRCMCR:			/* Coprocessor register transfer */
+			rpc = inpc + 12;
+		`DECODE_ALU:			/* ALU */
+			rpc = inpc + (insn[25] ? 8 : (insn[4] ? 12 : 8));
+		default:			/* X everything else out */
+			rpc = 32'hxxxxxxxx;
+		endcase
+	
+	always @(*) begin
+		read_0 = 4'hx;
+		read_1 = 4'hx;
+		read_2 = 4'hx;
+		
+		op0_out = 32'hxxxxxxxx;
+		op1_out = 32'hxxxxxxxx;
+		op2_out = 32'hxxxxxxxx;
+		carry_out = 1'bx;
+		
+		casez (insn)
+		`DECODE_ALU_MULT:	/* Multiply -- must come before ALU, because it pattern matches a specific case of ALU */
+		begin
+			read_0 = insn[15:12]; /* Rn */
+			read_1 = insn[3:0];   /* Rm */
+			read_2 = insn[11:8];  /* Rs */
+			
+			op0_out = regs0;
+			op1_out = regs1;
+			op2_out = regs2;
+		end
+//		`DECODE_ALU_MUL_LONG:	/* Multiply long */
+//		begin
+//			read_0 = insn[11:8]; /* Rn */
+//			read_1 = insn[3:0];   /* Rm */
+//			read_2 = 4'b0;       /* anyus */
+//
+//			op1_res = regs1;
+//		end
+		`DECODE_ALU_MRS:	/* MRS (Transfer PSR to register) */
+		begin end
+		`DECODE_ALU_MSR:	/* MSR (Transfer register to PSR) */
+		begin
+			read_0 = insn[3:0];	/* Rm */
+			
+			op0_out = regs0;
+		end
+		`DECODE_ALU_MSR_FLAGS:	/* MSR (Transfer register or immediate to PSR, flag bits only) */
+		begin
+			read_0 = insn[3:0];	/* Rm */
+			
+			if(insn[25]) begin     /* the constant case */
+				op0_out = rotate_res;
+			end else begin
+				op0_out = regs0;
+			end
+		end
+		`DECODE_ALU_SWP:	/* Atomic swap */
+		begin
+			read_0 = insn[19:16]; /* Rn */
+			read_1 = insn[3:0];   /* Rm */
+			
+			op0_out = regs0;
+			op1_out = regs1;
+		end
+		`DECODE_ALU_BX:		/* Branch and exchange */
+		begin
+			read_0 = insn[3:0];   /* Rn */
+			
+			op0_out = regs0;
+		end
+		`DECODE_ALU_HDATA_REG:	/* Halfword transfer - register offset */
+		begin
+			read_0 = insn[19:16];
+			read_1 = insn[3:0];
+			read_2 = insn[15:12];
 
-	assign regs0 = (regsel0 == 4'b1111) ? rpc : iregs0;
-	assign regs1 = (regsel1 == 4'b1111) ? rpc : iregs1;
-	assign regs2 = iregs2; /* use regs2 for things that cannot be r15 */
-
-	IHATEARMSHIFT(.insn(ansn),
-	              .operand(regs1),
-	              .reg_amt(regs2),
-	              .cflag_in(cps_in[`COND_CBIT]),
-	              .res(shift_res),
-	              .cflag_out(shift_cflag));
-
-	always @ (*) begin
-		casez (ansn)
-		32'b????000000??????????????1001????: begin /* Multiply */
-			rpc = inpc - 8;
-			regsel0 = ansn[15:12]; /* Rn */
-			regsel1 = ansn[3:0];   /* Rm */
-			regsel2 = ansn[11:8];  /* Rs */
-			op1_res = regs1;
-			new_cps = cps_in;
-		end
-/*
-		32'b????00001???????????????1001????: begin * Multiply long *
-			regsel0 = ansn[11:8]; * Rn *
-			regsel1 = ansn[3:0];  * Rm *
-			regsel2 = 4'b0;       * anyus *
-			op1_res = regs1;
-		end
-*/
-		32'b????00010?001111????000000000000: begin /* MRS (Transfer PSR to register) */
-			rpc = inpc - 8;
-			new_cps = cps_in;
-		end
-        	32'b????00010?101001111100000000????: begin /* MSR (Transfer register to PSR) */
-			rpc = inpc - 8;
-			new_cps = cps_in;
-        	end
-                32'b????00?10?1010001111????????????: begin /* MSR (Transfer register or immediate to PSR, flag bits onry) */
-			rpc = inpc - 8;
-			new_cps = cps_in;
-                end
-		32'b????00??????????????????????????: begin /* ALU */
-			rpc = inpc - (ansn[25] ? 8 : (ansn[4] ? 12 : 8));
-			regsel0 = ansn[19:16]; /* Rn */
-			regsel1 = ansn[3:0];   /* Rm */
-			regsel2 = ansn[11:8];  /* Rs for shift */
-			if(ansn[25]) begin     /* the constant case */
-				new_cps = cps_in;
-				op1_res = ({24'b0, ansn[7:0]} >> {ansn[11:8], 1'b0}) | ({24'b0, ansn[7:0]} << (5'b0 - {ansn[11:8], 1'b0}));
+			op0_out = regs0;
+			op1_out = regs1;
+			op2_out = regs2;
+		end
+		`DECODE_ALU_HDATA_IMM:	/* Halfword transfer - immediate offset */
+		begin
+			read_0 = insn[19:16];
+			read_1 = insn[15:12];
+			
+			op0_out = regs0;
+			op1_out = {24'b0, insn[11:8], insn[3:0]};
+			op2_out = regs1;
+		end
+		`DECODE_ALU:		/* ALU */
+		begin
+			read_0 = insn[19:16]; /* Rn */
+			read_1 = insn[3:0];   /* Rm */
+			read_2 = insn[11:8];  /* Rs for shift */
+			
+			op0_out = regs0;
+			if(insn[25]) begin     /* the constant case */
+				carry_out = incpsr[`CPSR_C];
+				op1_out = rotate_res;
 			end else begin
-				new_cps = {shift_cflag_out, cps_in[30:0]};
-				op1_res = shift_res;
+				carry_out = shift_cflag_out;
+				op1_out = shift_res;
 			end
 		end
-		32'b????00010?00????????00001001????: begin /* Atomic swap */
-			rpc = inpc - 8;
-			regsel0 = ansn[19:16]; /* Rn */
-			regsel1 = ansn[3:0];   /* Rm */
-			regsel2 = 4'b0;        /* anyus */
-			op1_res = regs1;
-		end
-		32'b????000100101111111111110001????: begin /* Branch and exchange */
-			rpc = inpc - 8;
-			regsel0 = ansn[3:0];   /* Rn */
-			new_cps = cps_in;
-		end
-		32'b????000??0??????????00001??1????: begin /* Halfword transfer - register offset */
-			rpc = inpc - 8;
-			regsel0 = ansn[19:16];
-			regsel1 = ansn[3:0];
-			regsel2 = 4'b0;
-			op1_res = regs1;
-			new_cps = cps_in;
-		end
-		32'b????000??1??????????00001??1????: begin /* Halfword transfer - immediate offset */
-			rpc = inpc - 8;
-			regsel0 = ansn[19:16];
-			regsel1 = ansn[3:0];
-			op1_res = {24'b0, ansn[11:8], ansn[3:0]};
-			new_cps = cps_in;
-		end
-		32'b????011????????????????????1????: begin /* Undefined. I hate ARM */
+		`DECODE_LDRSTR_UNDEFINED:	/* Undefined. I hate ARM */
+		begin
 			/* eat shit */
 		end
-		32'b????01??????????????????????????: begin /* Single data transfer */
-			rpc = inpc - 8;
-			regsel0 = ansn[19:16]; /* Rn */
-			regsel1 = ansn[3:0];   /* Rm */
-			if(ansn[25]) begin
-				op1_res = {20'b0, ansn[11:0]};
-				new_cps = cps_in;
+		`DECODE_LDRSTR:		/* Single data transfer */
+		begin
+			read_0 = insn[19:16]; /* Rn */
+			read_1 = insn[3:0];   /* Rm */
+			read_2 = insn[15:12];
+			
+			op0_out = regs0;
+			if(!insn[25] /* immediate */) begin
+				op1_out = {20'b0, insn[11:0]};
+				carry_out = incpsr[`CPSR_C];
 			end else begin
-				op1_res = shift_res;
-				new_cps = shift_cflag_out;
+				op1_out = shift_res;
+				carry_out = shift_cflag_out;
 			end
+			op2_out = regs2;
 		end
-		32'b????100?????????????????????????: begin /* Block data transfer */
-			rpc = inpc - 8;
-			regsel0 = ansn[19:16];
-			op1_res = {16'b0, ansn[15:0]};
-			new_cps = cps_in;
+		`DECODE_LDMSTM:		/* Block data transfer */
+		begin
+			read_0 = insn[19:16];
+			
+			op0_out = regs0;
+			op1_out = {16'b0, insn[15:0]};
 		end
-		32'b????101?????????????????????????: begin /* Branch */
-			rpc = inpc - 8;
-			op1_res = {6{ansn[23]}, ansn[23:0], 2'b0};
-			new_cps = cps_in;
+		`DECODE_BRANCH:		/* Branch */
+		begin
+			op0_out = {{6{insn[23]}}, insn[23:0], 2'b0};
 		end
-		32'b????110?????????????????????????: begin /* Coprocessor data transfer */
-			rpc = inpc - 8;
-			regsel0 = ansn[19:16];
-			op1_res = {24'b0, ansn[7:0]};
-			new_cps = cps_in;
+		`DECODE_LDCSTC:		/* Coprocessor data transfer */
+		begin
+			read_0 = insn[19:16];
+			
+			op0_out = regs0;
+			op1_out = {24'b0, insn[7:0]};
 		end
-		32'b????1110???????????????????0????: begin /* Coprocessor data op */
-			rpc = inpc - 8;
-			new_cps = cps_in;
+		`DECODE_CDP:		/* Coprocessor data op */
+		begin
 		end
-		32'b????1110???????????????????1????: begin /* Coprocessor register transfer */
-			rpc = inpc - 8;
-			new_cps = cps_in;
+		`DECODE_MRCMCR:		/* Coprocessor register transfer */
+		begin
+			read_0 = insn[15:12];
+			
+			op0_out = regs0;
 		end
-		32'b????1111????????????????????????: begin /* SWI */
-			rpc = inpc - 8;
-			new_cps = cps_in;
+		`DECODE_SWI:		/* SWI */
+		begin
 		end
 		default:
+			$display("Undecoded instruction");
 		endcase
 	end
 
+	
 	always @ (posedge clk) begin
-		op0 <= regs0;   /* Rn - always */
-		op1 <= op1_res; /* 'operand 2' - Rm */
-		op2 <= regs2;   /* thirdedge - Rs */
-		cps_out <= new_cps;
+		if (!stall)
+		begin
+			op0 <= op0_out;   /* Rn - always */
+			op1 <= op1_out; /* 'operand 2' - Rm */
+			op2 <= op2_out;   /* thirdedge - Rs */
+			carry <= carry_out;
+			outcpsr <= incpsr;
+			outspsr <= inspsr;
+		end
 	end
 
 endmodule
 
-module IHATEARMSHIFT(
+module IREALLYHATEARMSHIFT(
 	input [31:0] insn,
 	input [31:0] operand,
 	input [31:0] reg_amt,
 	input cflag_in,
-	output [31:0] res,
-	output cflag_out
+	output reg [31:0] res,
+	output reg cflag_out
 );
-	wire [1:0] shift_type;
 	wire [5:0] shift_amt;
-	wire elanus;
-
-	shift_type = insn[6:5];
-	if(insn[4]) begin
-		shift_amt = {|reg_amt[7:5], reg_amt[4:0]};
-		elanus = 1'b1;
-	end else begin
-		shift_amt = {insn[11:7] == 5'b0, insn[11:7]};
-		elanus = 1'b0;
-	end
+	reg is_arith, is_rot;
+	wire rshift_cout;
+	wire [31:0] rshift_res;
+
+	assign shift_amt = insn[4] ? {|reg_amt[7:5], reg_amt[4:0]}     /* reg-specified shift */
+	                           : {insn[11:7] == 5'b0, insn[11:7]}; /* immediate shift */
 
-	/* might want to write our own damn shifter that does arithmetic/logical efficiently and stuff */
-	always @ (*) begin
-		case (shift_type)
+	SuckLessShifter barrel(.oper(operand),
+	                       .carryin(cflag_in),
+	                       .amt(shift_amt),
+	                       .is_arith(is_arith),
+	                       .is_rot(is_rot),
+	                       .res(rshift_res),
+	                       .carryout(rshift_cout));
+
+	always @(*)
+		case (insn[6:5])
 		`SHIFT_LSL: begin
-			{cflag_out, res} = {cflag_in, operand} << {elanus & shift_amt[5], shift_amt[4:0]};
+			/* meaningless */
+			is_rot = 1'b0;
+			is_arith = 1'b0;
+		end
+		`SHIFT_LSR: begin
+			is_rot = 1'b0;
+			is_arith = 1'b0;
+		end
+		`SHIFT_ASR: begin
+			is_rot = 1'b0;
+			is_arith = 1'b1;
 		end
+		`SHIFT_ROR: begin
+			is_rot = 1'b1;
+			is_arith = 1'b0;
+		end
+		endcase
+
+	always @(*)
+		case (insn[6:5]) /* shift type */
+		`SHIFT_LSL:
+			{cflag_out, res} = {cflag_in, operand} << {insn[4] & shift_amt[5], shift_amt[4:0]};
 		`SHIFT_LSR: begin
-			{res, cflag_out} = {operand, cflag_in} >> shift_amt;
+			res = rshift_res;
+			cflag_out = rshift_cout;
 		end
 		`SHIFT_ASR: begin
-			{res, cflag_out} = {operand, cflag_in} >> shift_amt | (operand[31] ? ~(33'h1FFFFFFFF >> shift_amt) : 33'b0);
+			res = rshift_res;
+			cflag_out = rshift_cout;
 		end
 		`SHIFT_ROR: begin
-			if(!elanus && shift_amt[4:0] == 5'b0) begin /* RRX x.x */
+			if(!insn[4] && shift_amt[4:0] == 5'b0) begin /* RRX x.x */
 				res = {cflag_in, operand[31:1]};
 				cflag_out = operand[0];
-			end else if(!shift_amt) begin
-				res = operand;
-				cflag_out = cflag_in;
 			end else begin
-				res = operand >> shift_amt[4:0] | operand << (5'b0 - shift_amt[4:0]);
-				cflag_out = operand[shift_amt[4:0] - 5'b1];
+				res = rshift_res;
+				cflag_out = rshift_cout;
 			end
 		end
-	end
+		endcase
+endmodule
+
+module SuckLessShifter(
+	input [31:0] oper,
+	input carryin,
+	input [5:0] amt,
+	input is_arith,
+	input is_rot,
+	output wire [31:0] res,
+	output wire carryout
+);
+
+	wire [32:0] stage1, stage2, stage3, stage4, stage5;
+
+	wire pushbits = is_arith & oper[31];
+
+	/* do a barrel shift */
+	assign stage1 = amt[5] ? {is_rot ? oper : {32{pushbits}}, oper[31]} : {oper, carryin};
+	assign stage2 = amt[4] ? {is_rot ? stage1[16:1] : {16{pushbits}}, stage1[32:17], stage1[16]} : stage1;
+	assign stage3 = amt[3] ? {is_rot ? stage2[8:1] : {8{pushbits}}, stage2[32:9], stage2[8]} : stage2;
+	assign stage4 = amt[2] ? {is_rot ? stage3[4:1] : {4{pushbits}}, stage3[32:5], stage3[4]} : stage3;
+	assign stage5 = amt[1] ? {is_rot ? stage4[2:1] : {2{pushbits}}, stage4[32:3], stage4[2]} : stage4;
+	assign {res, carryout} = amt[0] ? {is_rot ? stage5[1] : pushbits, stage5[32:2], stage5[1]} : stage5;
+
 endmodule
+
+module SuckLessRotator(
+	input [31:0] oper,
+	input [3:0] amt,
+	output wire [31:0] res
+);
+
+	wire [31:0] stage1, stage2, stage3;
+	assign stage1 = amt[3] ? {oper[15:0], oper[31:16]} : oper;
+	assign stage2 = amt[2] ? {stage1[7:0], stage1[31:8]} : stage1;
+	assign stage3 = amt[1] ? {stage2[3:0], stage2[31:4]} : stage2;
+	assign res    = amt[0] ? {stage3[1:0], stage3[31:2]} : stage3;
+
+endmodule
+