X-Git-Url: http://git.joshuawise.com/firearm.git/blobdiff_plain/979f2bd7aca637e8e546ec2675bef82c353b9aea..ba1c1a27e6fb12882b59cdbd80147a4c12a675b4:/Memory.v?ds=inline

diff --git a/Memory.v b/Memory.v
index 04220ca..1dad548 100644
--- a/Memory.v
+++ b/Memory.v
@@ -1,9 +1,29 @@
 `include "ARM_Constants.v"
 
+`define SWP_READING	2'b01
+`define SWP_WRITING	2'b10
+
+`define LSRH_MEMIO	3'b001
+`define LSRH_BASEWB	3'b010
+`define LSRH_WBFLUSH	3'b100
+
+`define LSR_MEMIO	4'b0001
+`define LSR_STRB_WR	4'b0010
+`define LSR_BASEWB	4'b0100
+`define LSR_WBFLUSH	4'b1000
+
+`define LSM_SETUP	4'b0001
+`define LSM_MEMIO	4'b0010
+`define LSM_BASEWB	4'b0100
+`define LSM_WBFLUSH	4'b1000
+
+
 module Memory(
 	input clk,
 	input Nrst,
 
+	input flush,
+
 	/* bus interface */
 	output reg [31:0] busaddr,
 	output reg rd_req,
@@ -11,6 +31,7 @@ module Memory(
 	input rw_wait,
 	output reg [31:0] wr_data,
 	input [31:0] rd_data,
+	output reg [2:0] data_size,
 
 	/* regfile interface */
 	output reg [3:0] st_read,
@@ -20,6 +41,9 @@ module Memory(
 	output reg cp_req,
 	input cp_ack,
 	input cp_busy,
+	output reg cp_rnw,	/* 1 = read from CP, 0 = write to CP */
+	input [31:0] cp_read,
+	output reg [31:0] cp_write,
 	
 	/* stage inputs */
 	input inbubble,
@@ -28,6 +52,9 @@ module Memory(
 	input [31:0] op0,
 	input [31:0] op1,
 	input [31:0] op2,
+	input [31:0] spsr,
+	input [31:0] cpsr,
+	input cpsrup,
 	input write_reg,
 	input [3:0] write_num,
 	input [31:0] write_data,
@@ -39,24 +66,40 @@ module Memory(
 	output reg [31:0] outinsn,
 	output reg out_write_reg = 1'b0,
 	output reg [3:0] out_write_num = 4'bxxxx,
-	output reg [31:0] out_write_data = 32'hxxxxxxxx
+	output reg [31:0] out_write_data = 32'hxxxxxxxx,
+	output reg [31:0] outspsr = 32'hxxxxxxxx,
+	output reg [31:0] outcpsr = 32'hxxxxxxxx,
+	output reg outcpsrup = 1'hx
 	);
 
-	reg [31:0] addr, raddr, next_regdata;
+	reg [31:0] addr, raddr, prev_raddr, next_regdata, next_outcpsr;
+	reg next_outcpsrup;
+	reg [31:0] prevaddr;
 	reg [3:0] next_regsel, cur_reg, prev_reg;
-	reg next_writeback, next_notdone, next_inc_next;
+	reg next_writeback;
+
+	reg next_outbubble;	
+	reg next_write_reg;
+	reg [3:0] next_write_num;
+	reg [31:0] next_write_data;
+
+	reg [3:0] lsr_state = 4'b0001, next_lsr_state;
 	reg [31:0] align_s1, align_s2, align_rddata;
 
-	wire next_outbubble;	
-	wire next_write_reg;
-	wire [3:0] next_write_num;
-	wire [31:0] next_write_data;
+	reg [2:0] lsrh_state = 3'b001, next_lsrh_state;
+	reg [31:0] lsrh_rddata;
+	reg [15:0] lsrh_rddata_s1;
+	reg [7:0] lsrh_rddata_s2;
 
 	reg [15:0] regs, next_regs;
-	reg started = 1'b0, next_started;
+	reg [3:0] lsm_state = 4'b0001, next_lsm_state;
+	reg [5:0] offset, prev_offset, offset_sel;
 
-	reg notdone = 1'b0;
-	reg inc_next = 1'b0;
+	reg [31:0] swp_oldval, next_swp_oldval;
+	reg [1:0] swp_state = 2'b01, next_swp_state;
+	
+	reg do_rd_data_latch;
+	reg [31:0] rd_data_latch = 32'hxxxxxxxx;
 
 	always @(posedge clk)
 	begin
@@ -66,149 +109,572 @@ module Memory(
 		out_write_reg <= next_write_reg;
 		out_write_num <= next_write_num;
 		out_write_data <= next_write_data;
-		notdone <= next_notdone;
-		inc_next <= next_inc_next;
-		regs <= next_regs;
-		prev_reg <= cur_reg;
-		started <= next_started;
+		if (!rw_wait)
+			prev_offset <= offset;
+		prev_raddr <= raddr;
+		outcpsr <= next_outcpsr;
+		outspsr <= spsr;
+		outcpsrup <= next_outcpsrup;
+		swp_state <= next_swp_state;
+		lsm_state <= next_lsm_state;
+		lsr_state <= next_lsr_state;
+		lsrh_state <= next_lsrh_state;
+		if (do_rd_data_latch)
+			rd_data_latch <= rd_data;
+		swp_oldval <= next_swp_oldval;
+		prevaddr <= addr;
 	end
-
+	
+	reg delayedflush = 0;
+	always @(posedge clk)
+		if (flush && outstall /* halp! I can't do it now, maybe later? */)
+			delayedflush <= 1;
+		else if (!outstall /* anything has been handled this time around */)
+			delayedflush <= 0;
+	
+	/* Drive the state machines and stall. */
 	always @(*)
 	begin
-		addr = 32'hxxxxxxxx;
-		raddr = 32'hxxxxxxxx;
-		rd_req = 1'b0;
-		wr_req = 1'b0;
-		wr_data = 32'hxxxxxxxx;
-		busaddr = 32'hxxxxxxxx;
-		outstall = 1'b0;
-		next_notdone = 1'b0;
-		next_write_reg = write_reg;
-		next_write_num = write_num;
-		next_write_data = write_data;
-		next_inc_next = 1'b0;
-		next_outbubble = inbubble;
 		outstall = 1'b0;
-		next_regs = 16'b0;
-		next_started = started;
-		cp_req = 1'b0;
-
+		next_lsm_state = lsm_state;
+		next_lsr_state = lsr_state;
+		next_lsrh_state = lsrh_state;
+		next_swp_state = swp_state;
 		casez(insn)
-		`DECODE_LDRSTR_UNDEFINED: begin end
-		`DECODE_LDRSTR: begin
-			if (!inbubble) begin
-				next_outbubble = rw_wait;
-				outstall = rw_wait | notdone;
-			
-				addr = insn[23] ? op0 + op1 : op0 - op1; /* up/down select */
-				raddr = insn[24] ? op0 : addr; /* pre/post increment */
-				busaddr = {raddr[31:2], 2'b0};
-				rd_req = insn[20];
-				wr_req = ~insn[20];
+		`DECODE_ALU_SWP: if(!inbubble) begin
+			case(swp_state)
+			`SWP_READING: begin
+				outstall = 1'b1;
+				if (!rw_wait)
+					next_swp_state = `SWP_WRITING;
+				$display("SWP: read stage");
+			end
+			`SWP_WRITING: begin
+				outstall = rw_wait;
+				if(!rw_wait)
+					next_swp_state = `SWP_READING;
+				$display("SWP: write stage");
+			end
+			default: begin
+				outstall = 1'bx;
+				next_swp_state = 2'bxx;
+			end
+			endcase
+		end
+		`DECODE_ALU_MULT: begin
+			outstall = 1'b0;	/* XXX work around for Xilinx bug */
+			next_lsrh_state = lsrh_state;
+		end
+		`DECODE_ALU_HDATA_REG,
+		`DECODE_ALU_HDATA_IMM: if(!inbubble) begin
+			case(lsrh_state)
+			`LSRH_MEMIO: begin
+				outstall = rw_wait;
+				if(insn[21] | !insn[24]) begin
+					outstall = 1'b1;
+					if(!rw_wait)
+						next_lsrh_state = `LSRH_BASEWB;
+				end
 				
-				/* rotate to correct position */
-				align_s1 = raddr[1] ? {rd_data[15:0], rd_data[31:16]} : rd_data;
-				align_s2 = raddr[0] ? {align_s1[7:0], align_s1[31:8]} : align_s1;
-				/* select byte or word */
-				align_rddata = insn[22] ? {24'b0, align_s2[7:0]} : align_s2;
+				if (flush) /* special case! */ begin
+					outstall = 1'b0;
+					next_lsrh_state = `LSRH_MEMIO;
+				end
 				
-				if(!insn[20]) begin
-					wr_data = insn[22] ? {4{op2[7:0]}} : op2; /* XXX need to actually store just a byte */
+				$display("ALU_LDRSTRH: rd_req %d, wr_req %d", rd_req, wr_req);
+			end
+			`LSRH_BASEWB: begin
+				outstall = 1'b1;
+				next_lsrh_state = `LSRH_WBFLUSH;
+			end
+			`LSRH_WBFLUSH: begin
+				outstall = 1'b0;
+				next_lsrh_state = `LSRH_MEMIO;
+			end
+			default: begin
+				outstall = 1'bx;
+				next_lsrh_state = 3'bxxx;
+			end
+			endcase
+		end
+		`DECODE_LDRSTR_UNDEFINED: begin end
+		`DECODE_LDRSTR: if(!inbubble) begin
+			outstall = rw_wait;
+			case(lsr_state)
+			`LSR_MEMIO: begin
+				outstall = rw_wait;
+				next_lsr_state = `LSR_MEMIO;
+				if (insn[22] /* B */ && !insn[20] /* L */) begin	/* i.e., strb */
+					outstall = 1'b1;
+					if (!rw_wait)
+						next_lsr_state = `LSR_STRB_WR;
+				end else if (insn[21] /* W */ || !insn[24] /* P */) begin	/* writeback needed */
+					outstall = 1'b1;
+					if (!rw_wait)
+						next_lsr_state = `LSR_BASEWB;
 				end
-				else if(!inc_next) begin
-					next_write_reg = 1'b1;
-					next_write_num = insn[15:12];
-					next_write_data = align_rddata;
-					next_inc_next = 1'b1;
+				
+				if (flush) begin
+					outstall = 1'b0;
+					next_lsr_state = `LSR_MEMIO;
+				end
+				$display("LDRSTR: rd_req %d, wr_req %d, raddr %08x, wait %d", rd_req, wr_req, raddr, rw_wait);
+			end
+			`LSR_STRB_WR: begin
+				outstall = 1;
+				if(insn[21] /* W */ | !insn[24] /* P */) begin
+					if(!rw_wait)
+						next_lsr_state = `LSR_BASEWB;
+				end else if (!rw_wait)
+					next_lsr_state = `LSR_WBFLUSH;
+				$display("LDRSTR: Handling STRB");
+			end
+			`LSR_BASEWB: begin
+				outstall = 1;
+				next_lsr_state = `LSR_WBFLUSH;
+			end
+			`LSR_WBFLUSH: begin
+				outstall = 0;
+				next_lsr_state = `LSR_MEMIO;
+			end
+			default: begin
+				outstall = 1'bx;
+				next_lsr_state = 4'bxxxx;
+			end
+			endcase
+			$display("LDRSTR: Decoded, bubble %d, insn %08x, lsm state %b -> %b, stall %d", inbubble, insn, lsr_state, next_lsr_state, outstall);
+		end
+		`DECODE_LDMSTM: if(!inbubble) begin
+			outstall = rw_wait;
+			case(lsm_state)
+			`LSM_SETUP: begin
+				outstall = 1'b1;
+				next_lsm_state = `LSM_MEMIO;
+				if (flush) begin
+					outstall = 1'b0;
+					next_lsm_state = `LSM_SETUP;
+				end
+				$display("LDMSTM: Round 1: base register: %08x, reg list %b", op0, op1[15:0]);
+			end
+			`LSM_MEMIO: begin
+				outstall = 1'b1;
+				if(next_regs == 16'b0 && !rw_wait) begin
+					next_lsm_state = `LSM_BASEWB;
 				end
-				else if(insn[21]) begin
+				
+				$display("LDMSTM: Stage 2: Writing: regs %b, next_regs %b, reg %d, wr_data %08x, addr %08x", regs, next_regs, cur_reg, st_data, busaddr);
+			end
+			`LSM_BASEWB: begin
+				outstall = 1;
+				next_lsm_state = `LSM_WBFLUSH;
+				$display("LDMSTM: Stage 3: Writing back");
+			end
+			`LSM_WBFLUSH: begin
+				outstall = 0;
+				next_lsm_state = `LSM_SETUP;
+			end
+			default: begin
+				outstall = 1'bx;
+				next_lsm_state = 4'bxxxx;
+			end
+			endcase
+			$display("LDMSTM: Decoded, bubble %d, insn %08x, lsm state %b -> %b, stall %d", inbubble, insn, lsm_state, next_lsm_state, outstall);
+		end
+		`DECODE_LDCSTC: if(!inbubble) begin
+			$display("WARNING: Unimplemented LDCSTC");
+		end
+		`DECODE_CDP: if (!inbubble) begin
+			if (cp_busy) begin
+				outstall = 1;
+			end
+			if (!cp_ack) begin
+				/* XXX undefined instruction trap */
+				$display("WARNING: Possible CDP undefined instruction");
+			end
+		end
+		`DECODE_MRCMCR: if (!inbubble) begin
+			if (cp_busy) begin
+				outstall = 1;
+			end
+			if (!cp_ack) begin
+				$display("WARNING: Possible MRCMCR undefined instruction: cp_ack %d, cp_busy %d",cp_ack, cp_busy);
+			end
+			$display("MRCMCR: ack %d, busy %d", cp_ack, cp_busy);
+		end
+		default: begin end
+		endcase
+	end
+	
+	/* Coprocessor input. */
+	always @(*)
+	begin
+		cp_req = 0;
+		cp_rnw = 1'bx;
+		cp_write = 32'hxxxxxxxx;
+		casez (insn)
+		`DECODE_CDP: if(!inbubble) begin
+			cp_req = 1;
+		end
+		`DECODE_MRCMCR: if(!inbubble) begin
+			cp_req = 1;
+			cp_rnw = insn[20] /* L */;
+			if (insn[20] == 0 /* store to coprocessor */)
+				cp_write = op0;
+		end
+		endcase
+	end
+	
+	/* Register output logic. */
+	always @(*)
+	begin
+		next_write_reg = write_reg;
+		next_write_num = write_num;
+		next_write_data = write_data;
+		next_outcpsr = lsm_state == 4'b0010 ? outcpsr : cpsr;
+		next_outcpsrup = cpsrup;
+		
+		casez(insn)
+		`DECODE_ALU_SWP: if (!inbubble) begin
+			next_write_reg = 1'bx;
+			next_write_num = 4'bxxxx;
+			next_write_data = 32'hxxxxxxxx;
+			case(swp_state)
+			`SWP_READING:
+				next_write_reg = 1'b0;
+			`SWP_WRITING: begin
+				next_write_reg = 1'b1;
+				next_write_num = insn[15:12];
+				next_write_data = insn[22] ? {24'b0, swp_oldval[7:0]} : swp_oldval;
+			end
+			default: begin end
+			endcase
+		end
+		`DECODE_ALU_MULT: begin
+			next_write_reg = write_reg;	/* XXX workaround for ISE 10.1 bug */
+			next_write_num = write_num;
+			next_write_data = write_data;
+			next_outcpsr = lsm_state == 4'b0010 ? outcpsr : cpsr;
+			next_outcpsrup = cpsrup;
+		end
+		`DECODE_ALU_HDATA_REG,
+		`DECODE_ALU_HDATA_IMM: if(!inbubble) begin
+			next_write_reg = 1'bx;
+			next_write_num = 4'bxxxx;
+			next_write_data = 32'hxxxxxxxx;
+			case(lsrh_state)
+			`LSRH_MEMIO: begin
+				next_write_num = insn[15:12];
+				next_write_data = lsrh_rddata;
+				if(insn[20]) begin
 					next_write_reg = 1'b1;
-					next_write_num = insn[19:16];
-					next_write_data = addr;
 				end
-				next_notdone = rw_wait & insn[20] & insn[21];
 			end
+			`LSRH_BASEWB: begin
+				next_write_reg = 1'b1;
+				next_write_num = insn[19:16];
+				next_write_data = addr;
+			end
+			`LSRH_WBFLUSH:
+				next_write_reg = 1'b0;
+			default: begin end
+			endcase
 		end
-		`DECODE_LDMSTM: begin
-			rd_req = insn[20];
-			wr_req = ~insn[20];
-			if(!started) begin
-				next_regs = op1[15:0];
-				next_started = 1'b1;
+		`DECODE_LDRSTR_UNDEFINED: begin end
+		`DECODE_LDRSTR: if(!inbubble) begin
+			next_write_reg = 1'bx;
+			next_write_num = 4'bxxxx;
+			next_write_data = 32'hxxxxxxxx;
+			case(lsr_state)
+			`LSR_MEMIO: begin
+				next_write_reg = insn[20] /* L */;
+				next_write_num = insn[15:12];
+				if(insn[20] /* L */) begin
+					next_write_data = insn[22] /* B */ ? {24'h0, align_rddata[7:0]} : align_rddata;
+				end
+			end
+			`LSR_STRB_WR:
+				next_write_reg = 1'b0;
+			`LSR_BASEWB: begin
+				next_write_reg = 1'b1;
+				next_write_num = insn[19:16];
+				next_write_data = addr;
+			end
+			`LSR_WBFLUSH:
+				next_write_reg = 1'b0;
+			default: begin end
+			endcase
+		end
+		`DECODE_LDMSTM: if(!inbubble) begin
+			next_write_reg = 1'bx;
+			next_write_num = 4'bxxxx;
+			next_write_data = 32'hxxxxxxxx;
+			case(lsm_state)
+			`LSM_SETUP:
+				next_write_reg = 1'b0;
+			`LSM_MEMIO: begin
+				if(insn[20] /* L */) begin
+					next_write_reg = !rw_wait;
+					next_write_num = cur_reg;
+					next_write_data = rd_data;
+				end else
+					next_write_reg = 1'b0;
+			end
+			`LSM_BASEWB: begin
+				next_write_reg = insn[21] /* writeback */;
+				next_write_num = insn[19:16];
+				next_write_data = insn[23] ? op0 + {26'b0, prev_offset} : op0 - {26'b0, prev_offset};
+				if(cur_reg == 4'hF && insn[22]) begin
+					next_outcpsr = spsr;
+					next_outcpsrup = 1;
+				end
 			end
-			else if(inc_next) begin
-				if(insn[21]) begin
+			`LSM_WBFLUSH:
+				next_write_reg = 1'b0;
+			default: begin end
+			endcase
+		end
+		`DECODE_MRCMCR: if(!inbubble) begin
+			next_write_reg = 1'bx;
+			next_write_num = 4'bxxxx;
+			next_write_data = 32'hxxxxxxxx;
+			next_outcpsr = 32'hxxxxxxxx;
+			next_outcpsrup = 1'bx;
+			if (insn[20] == 1 /* load from coprocessor */)
+				if (insn[15:12] != 4'hF /* Fuck you ARM */) begin
 					next_write_reg = 1'b1;
-					next_write_num = insn[19:16];
-					next_write_data = op0;
+					next_write_num = insn[15:12];
+					next_write_data = cp_read;
+				end else begin
+					next_outcpsr = {cp_read[31:28], cpsr[27:0]};
+					next_outcpsrup = 1;
 				end
-				next_started = 1'b0;
+		end
+		endcase
+	end
+	
+	/* Bus/address control logic. */
+	always @(*)
+	begin
+		rd_req = 1'b0;
+		wr_req = 1'b0;
+		offset = prev_offset;
+		addr = prevaddr;
+		raddr = 32'hxxxxxxxx;
+		busaddr = 32'hxxxxxxxx;
+		data_size = 3'bxxx;
+		
+		casez(insn)
+		`DECODE_ALU_SWP: if(!inbubble) begin
+			busaddr = {op0[31:2], 2'b0};
+			data_size = insn[22] ? 3'b001 : 3'b100;
+			case(swp_state)
+			`SWP_READING:
+				rd_req = 1'b1;
+			`SWP_WRITING:
+				wr_req = 1'b1;
+			default: begin end
+			endcase
+		end
+		`DECODE_ALU_MULT: begin
+			rd_req = 1'b0;	/* XXX workaround for Xilinx bug */
+			wr_req = 1'b0;
+			offset = prev_offset;
+			addr = prevaddr;
+		end
+		`DECODE_ALU_HDATA_REG,
+		`DECODE_ALU_HDATA_IMM: if(!inbubble) begin
+			addr = insn[23] ? op0 + op1 : op0 - op1; /* up/down select */
+			raddr = insn[24] ? op0 : addr; /* pre/post increment */
+			busaddr = raddr;
+			/* rotate to correct position */
+			case(insn[6:5])
+			2'b01: /* unsigned half */
+				data_size = 3'b010;
+			2'b10: /* signed byte */
+				data_size = 3'b001;
+			2'b11: /* signed half */
+				data_size = 3'b010;
+			default: begin
+				data_size = 3'bxxx;
 			end
-			else if(rw_wait) begin
-				next_regs = regs;
-				cur_reg = prev_reg;
+			endcase
+			
+			case(lsrh_state)
+			`LSRH_MEMIO: begin
+				rd_req = insn[20];
+				wr_req = ~insn[20];
+			end
+			`LSRH_BASEWB: begin end
+			`LSRH_WBFLUSH: begin end
+			default: begin end
+			endcase
+		end
+		`DECODE_LDRSTR_UNDEFINED: begin end
+		`DECODE_LDRSTR: if(!inbubble) begin
+			addr = insn[23] ? op0 + op1 : op0 - op1; /* up/down select */
+			raddr = insn[24] ? addr : op0; /* pre/post increment */
+			busaddr = raddr;
+			data_size = insn[22] ? 3'b001 : 3'b100;
+			case (lsr_state)
+			`LSR_MEMIO: begin
+				rd_req = insn[20] /* L */ || insn[22] /* B */;
+				wr_req = !insn[20] /* L */ && !insn[22]/* B */;
 			end
-			else begin
+			`LSR_STRB_WR:
+				wr_req = 1;
+			`LSR_BASEWB: begin end
+			`LSR_WBFLUSH: begin end
+			default: begin end
+			endcase
+		end
+		`DECODE_LDMSTM: if (!inbubble) begin
+			data_size = 3'b100;
+			case (lsm_state)
+			`LSM_SETUP:
+				offset = 6'b0;
+			`LSM_MEMIO: begin
+				rd_req = insn[20];
+				wr_req = ~insn[20];
+				offset = prev_offset + 6'h4;
+				offset_sel = insn[24] ? offset : prev_offset;
+				raddr = insn[23] ? op0 + {26'b0, offset_sel} : op0 - {26'b0, offset_sel};
+				busaddr = raddr;
+			end
+			`LSM_BASEWB: begin end
+			`LSM_WBFLUSH: begin end
+			default: begin end
+			endcase
+		end
+		`DECODE_LDCSTC: begin end
+		`DECODE_CDP: begin end
+		`DECODE_MRCMCR: begin end
+		default: begin end
+		endcase
+	end
+	
+	/* Bus data control logic. */
+	always @(*)
+	begin
+		wr_data = 32'hxxxxxxxx;
+		
+		casez(insn)
+		`DECODE_ALU_SWP: if(!inbubble)
+			if (swp_state == `SWP_WRITING)
+				wr_data = insn[22] ? {4{op1[7:0]}} : op1;
+		`DECODE_ALU_MULT: begin end
+		`DECODE_ALU_HDATA_REG,
+		`DECODE_ALU_HDATA_IMM: if(!inbubble)
+			case(insn[6:5])
+			2'b01: /* unsigned half */
+				wr_data = {2{op2[15:0]}}; /* XXX need to store halfword */
+			2'b10: /* signed byte */
+				wr_data = {4{op2[7:0]}};
+			2'b11: /* signed half */
+				wr_data = {2{op2[15:0]}};
+			default: begin end
+			endcase
+		`DECODE_LDRSTR_UNDEFINED: begin end
+		`DECODE_LDRSTR: if(!inbubble) begin
+			wr_data = insn[22] ? {24'h0, {op2[7:0]}} : op2;
+			if (lsr_state == `LSR_STRB_WR)
+				case (busaddr[1:0])
+				2'b00: wr_data = {rd_data_latch[31:8], op2[7:0]};
+				2'b01: wr_data = {rd_data_latch[31:16], op2[7:0], rd_data_latch[7:0]};
+				2'b10: wr_data = {rd_data_latch[31:24], op2[7:0], rd_data_latch[15:0]};
+				2'b11: wr_data = {op2[7:0], rd_data_latch[23:0]};
+				endcase
+		end
+		`DECODE_LDMSTM: if (!inbubble)
+			if (lsm_state == `LSM_MEMIO)
+				wr_data = (cur_reg == 4'hF) ? (pc + 12) : st_data;
+		`DECODE_LDCSTC: begin end
+		`DECODE_CDP: begin end
+		`DECODE_MRCMCR: begin end
+		default: begin end
+		endcase
+	end
+	
+	/* LDM/STM register control logic. */
+	always @(posedge clk)
+		if (!rw_wait || lsm_state != `LSM_MEMIO)
+		begin
+			prev_reg <= cur_reg;
+			regs <= next_regs;
+		end
+	
+	always @(*)
+	begin
+		st_read = 4'hx;
+		cur_reg = prev_reg;
+		next_regs = regs;
+		
+		casez(insn)
+		`DECODE_LDMSTM: if(!inbubble) begin
+			case(lsm_state)
+			`LSM_SETUP:
+				next_regs = insn[23] /* U */ ? op1[15:0] : {op1[0], op1[1], op1[2], op1[3], op1[4], op1[5], op1[6], op1[7],
+				                                            op1[8], op1[9], op1[10], op1[11], op1[12], op1[13], op1[14], op1[15]};
+			`LSM_MEMIO: begin
 				casez(regs)
 				16'b???????????????1: begin
 					cur_reg = 4'h0;
-					next_regs = regs & 16'b1111111111111110;
+					next_regs = {regs[15:1], 1'b0};
 				end
 				16'b??????????????10: begin
 					cur_reg = 4'h1;
-					next_regs = regs & 16'b1111111111111100;
+					next_regs = {regs[15:2], 2'b0};
 				end
 				16'b?????????????100: begin
 					cur_reg = 4'h2;
-					next_regs = regs & 16'b1111111111111000;
+					next_regs = {regs[15:3], 3'b0};
 				end
 				16'b????????????1000: begin
 					cur_reg = 4'h3;
-					next_regs = regs & 16'b1111111111110000;
+					next_regs = {regs[15:4], 4'b0};
 				end
 				16'b???????????10000: begin
 					cur_reg = 4'h4;
-					next_regs = regs & 16'b1111111111100000;
+					next_regs = {regs[15:5], 5'b0};
 				end
 				16'b??????????100000: begin
 					cur_reg = 4'h5;
-					next_regs = regs & 16'b1111111111000000;
+					next_regs = {regs[15:6], 6'b0};
 				end
 				16'b?????????1000000: begin
 					cur_reg = 4'h6;
-					next_regs = regs & 16'b1111111110000000;
+					next_regs = {regs[15:7], 7'b0};
 				end
 				16'b????????10000000: begin
 					cur_reg = 4'h7;
-					next_regs = regs & 16'b1111111100000000;
+					next_regs = {regs[15:8], 8'b0};
 				end
 				16'b???????100000000: begin
 					cur_reg = 4'h8;
-					next_regs = regs & 16'b1111111000000000;
+					next_regs = {regs[15:9], 9'b0};
 				end
 				16'b??????1000000000: begin
 					cur_reg = 4'h9;
-					next_regs = regs & 16'b1111110000000000;
+					next_regs = {regs[15:10], 10'b0};
 				end
 				16'b?????10000000000: begin
 					cur_reg = 4'hA;
-					next_regs = regs & 16'b1111100000000000;
+					next_regs = {regs[15:11], 11'b0};
 				end
 				16'b????100000000000: begin
 					cur_reg = 4'hB;
-					next_regs = regs & 16'b1111000000000000;
+					next_regs = {regs[15:12], 12'b0};
 				end
 				16'b???1000000000000: begin
 					cur_reg = 4'hC;
-					next_regs = regs & 16'b1110000000000000;
+					next_regs = {regs[15:13], 13'b0};
 				end
 				16'b??10000000000000: begin
 					cur_reg = 4'hD;
-					next_regs = regs & 16'b1100000000000000;
+					next_regs = {regs[15:14], 14'b0};
 				end
 				16'b?100000000000000: begin
 					cur_reg = 4'hE;
-					next_regs = regs & 16'b1000000000000000;
+					next_regs = {regs[15], 15'b0};
 				end
 				16'b1000000000000000: begin
 					cur_reg = 4'hF;
@@ -219,11 +685,124 @@ module Memory(
 					next_regs = 16'b0;
 				end
 				endcase
-				next_inc_next = next_regs == 16'b0;
-				next_notdone = ~next_inc_next | (rw_wait & insn[20] & insn[21]);
+				cur_reg = insn[23] ? cur_reg : 4'hF - cur_reg;
+				
+				st_read = cur_reg;
+			end
+			`LSM_BASEWB: begin end
+			`LSM_WBFLUSH: begin end
+			default: begin end
+			endcase
+		end
+		endcase
+	end
+	
+	always @(*)
+	begin
+		do_rd_data_latch = 0;
+		
+		next_outbubble = inbubble;
+		
+		lsrh_rddata = 32'hxxxxxxxx;
+		lsrh_rddata_s1 = 16'hxxxx;
+		lsrh_rddata_s2 = 8'hxx;
+		next_swp_oldval = swp_oldval;
+		
+		align_s1 = 32'hxxxxxxxx;
+		align_s2 = 32'hxxxxxxxx;
+		align_rddata = 32'hxxxxxxxx;
+
+		/* XXX shit not given about endianness */
+		casez(insn)
+		`DECODE_ALU_SWP: if(!inbubble) begin
+			next_outbubble = rw_wait;
+			case(swp_state)
+			`SWP_READING:
+				if(!rw_wait)
+					next_swp_oldval = rd_data;
+			`SWP_WRITING: begin end
+			default: begin end
+			endcase
+		end
+		`DECODE_ALU_MULT: begin
+			next_outbubble = inbubble;	/* XXX workaround for Xilinx bug */
+		end
+		`DECODE_ALU_HDATA_REG,
+		`DECODE_ALU_HDATA_IMM: if(!inbubble) begin
+			next_outbubble = rw_wait;
+			
+			/* rotate to correct position */
+			case(insn[6:5])
+			2'b01: begin /* unsigned half */
+				lsrh_rddata = {16'b0, raddr[1] ? rd_data[31:16] : rd_data[15:0]};
+			end
+			2'b10: begin /* signed byte */
+				lsrh_rddata_s1 = raddr[1] ? rd_data[31:16] : rd_data[15:0];
+				lsrh_rddata_s2 = raddr[0] ? lsrh_rddata_s1[15:8] : lsrh_rddata_s1[7:0];
+				lsrh_rddata = {{24{lsrh_rddata_s2[7]}}, lsrh_rddata_s2};
+			end
+			2'b11: begin /* signed half */
+				lsrh_rddata = raddr[1] ? {{16{rd_data[31]}}, rd_data[31:16]} : {{16{rd_data[15]}}, rd_data[15:0]};
+			end
+			default: begin
+				lsrh_rddata = 32'hxxxxxxxx;
+			end
+			endcase
+
+			case(lsrh_state)
+			`LSRH_MEMIO: begin end
+			`LSRH_BASEWB:
+				next_outbubble = 1'b0;
+			`LSRH_WBFLUSH: begin end
+			default: begin end
+			endcase
+		end
+		`DECODE_LDRSTR_UNDEFINED: begin end
+		`DECODE_LDRSTR: if(!inbubble) begin
+			next_outbubble = rw_wait;
+			/* rotate to correct position */
+			align_s1 = raddr[1] ? {rd_data[15:0], rd_data[31:16]} : rd_data;
+			align_s2 = raddr[0] ? {align_s1[7:0], align_s1[31:8]} : align_s1;
+			/* select byte or word */
+			align_rddata = insn[22] ? {24'b0, align_s2[7:0]} : align_s2;
+			case(lsr_state)
+			`LSR_MEMIO:
+				if (insn[22] /* B */ && !insn[20] /* L */)
+					do_rd_data_latch = 1;
+			`LSR_STRB_WR: begin end
+			`LSR_BASEWB:
+				next_outbubble = 0;
+			`LSR_WBFLUSH: begin end
+			default: begin end
+			endcase
+		end
+		/* XXX ldm/stm incorrect in that stupid case where one of the listed regs is the base reg */
+		`DECODE_LDMSTM: if(!inbubble) begin
+			next_outbubble = rw_wait;
+			case(lsm_state)
+			`LSM_SETUP: begin end
+			`LSM_MEMIO: begin end
+			`LSM_BASEWB:
+				next_outbubble = 0;
+			`LSM_WBFLUSH: begin end
+			default: $stop;
+			endcase
+		end
+		`DECODE_LDCSTC: begin end
+		`DECODE_CDP: if(!inbubble) begin
+			if (cp_busy) begin
+				next_outbubble = 1;
+			end
+		end
+		`DECODE_MRCMCR: if(!inbubble) begin
+			if (cp_busy) begin
+				next_outbubble = 1;
 			end
 		end
 		default: begin end
 		endcase
+		
+		if ((flush || delayedflush) && !outstall)
+			next_outbubble = 1'b1;
 	end
 endmodule