[firearm.git] / Issue.v

`include "ARM_Constants.v"

module Issue(
	input clk,
	input Nrst,	/* XXX not used yet */
	
	input stall_1a,	/* pipeline control */
	input flush,	/* XXX not used yet */
	
	input bubble_1a,	/* stage inputs */
	input [31:0] insn_1a,
	input [31:0] pc_1a,
	input [31:0] cpsr_1a,
	
	output wire stall_0a,	/* stage outputs */
	output reg bubble_2a = 1,
	output reg [31:0] pc_2a = 0,
	output reg [31:0] insn_2a = 0
	/* XXX other? */
	);
	
`ifdef COPY_PASTA_FODDER
	/* from page 2 of ARM7TDMIvE2.pdf */
	casex (insn_1a)
	`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 */
	`DECODE_ALU_HDATA_REG:	/* Halfword transfer - register offset */
	`DECODE_ALU_HDATA_IMM:	/* Halfword transfer - immediate offset */
	`DECODE_ALU:		/* ALU */
	`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_MRCMCR:		/* Coprocessor register transfer */
	`DECODE_SWI:		/* SWI */
	default:		/* X everything else out */
	endcase
`endif

	/* Flag setting */
	reg use_cpsr;
	reg [15:0] use_regs;
	reg def_cpsr;
	reg [15:0] def_regs;
	
	function [15:0] idxbit;
		input [3:0] r;
		if (r == 15)
			idxbit = 0;
		else
			idxbit = (16'b1) << r;
	endfunction
	
	wire [3:0] rn = insn_1a[19:16];
	wire [3:0] rd = insn_1a[15:12];
	wire [3:0] rs = insn_1a[11:8];
	wire [3:0] rm = insn_1a[3:0];
	wire [3:0] cond = insn_1a[31:28];
	
	wire [3:0] rd_mul = insn_1a[19:16];
	wire [3:0] rn_mul = insn_1a[15:12];
	wire [3:0] rs_mul = insn_1a[11:8];
	
	wire [3:0] alu_opc = insn_1a[24:21];
	
	function alu_is_logical;
		input [3:0] op;
		
		case (op)
		`ALU_AND,`ALU_EOR,`ALU_TST,`ALU_TEQ,`ALU_ORR,`ALU_MOV,`ALU_BIC,`ALU_MVN: alu_is_logical = 1;
		default: alu_is_logical = 0;
		endcase
	endfunction
	
	function alu_flags_only;
		input [3:0] op;
		
		case (op)
		`ALU_TST,`ALU_TEQ,`ALU_CMP,`ALU_CMN: alu_flags_only = 1;
		default: alu_flags_only = 0;
		endcase
	endfunction
	
	function shift_requires_carry;
		input [7:0] shift;
		
		case(shift[1:0])
		`SHIFT_LSL: shift_requires_carry = (shift[7:2] == 0);
		`SHIFT_LSR: shift_requires_carry = 0;
		`SHIFT_ASR: shift_requires_carry = 0;
		`SHIFT_ROR: shift_requires_carry = (shift[7:2] == 0);
		endcase
	endfunction
	
	always @(*)
		casez (insn_1a)
		`DECODE_ALU_MULT:	/* Multiply -- must come before ALU, because it pattern matches a specific case of ALU */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = (insn_1a[21] /* accum */ ? idxbit(rn_mul) : 0) | idxbit(rs_mul) | idxbit(rm);
			def_cpsr = insn_1a[20] /* setcc */;
			def_regs = idxbit(rd_mul);
		end
//		`DECODE_ALU_MUL_LONG:	/* Multiply long */
		`DECODE_ALU_MRS:	/* MRS (Transfer PSR to register) */
		begin
			use_cpsr = `COND_MATTERS(cond) || (insn_1a[22] == 0) /* Source = CPSR */;
			use_regs = 0;
			def_cpsr = 0;
			def_regs = idxbit(rd);
		end
		`DECODE_ALU_MSR:	/* MSR (Transfer register to PSR) */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rm);
			def_cpsr = 1;
			def_regs = 0;
		end
		`DECODE_ALU_MSR_FLAGS:	/* MSR (Transfer register or immediate to PSR, flag bits only) */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = insn_1a[25] ? 0 : idxbit(rm);
			def_cpsr = 1;
			def_regs = 0;
		end
		`DECODE_ALU_SWP:	/* Atomic swap */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rn) | idxbit(rm);
			def_cpsr = 0;
			def_regs = idxbit(rd);
		end
		`DECODE_ALU_BX:	/* Branch */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rm);
			def_cpsr = 0;	// don't care, we'll never get there
			def_regs = 0;
		end
		`DECODE_ALU_HDATA_REG:	/* Halfword transfer - register offset */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rn) | idxbit(rm) | (insn_1a[20] /* L */ ? 0 : idxbit(rd));
			def_cpsr = 0;
			def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
		end
		`DECODE_ALU_HDATA_IMM:	/* Halfword transfer - immediate offset */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rn) | (insn_1a[20] /* L */ ? 0 : idxbit(rd));
			def_cpsr = 0;
			def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
		end
		`DECODE_ALU:	/* ALU */
		begin
			use_cpsr = `COND_MATTERS(cond) | (!insn_1a[25] /* I */ && shift_requires_carry(insn_1a[11:4]));
			use_regs =
				(insn_1a[25] /* I */ ? 0 :
					(insn_1a[4] /* shift by reg */ ?
						(idxbit(rs) | idxbit(rm)) :
						(idxbit(rm)))) |
				(((alu_opc != `ALU_MOV) && (alu_opc != `ALU_MVN)) ? idxbit(rn) : 0);
			def_cpsr = insn_1a[20] /* S */;
			def_regs = alu_flags_only(alu_opc) ? 0 : idxbit(rd);
		end
		`DECODE_LDRSTR_UNDEFINED:	/* Undefined. I hate ARM */
		begin	
			use_cpsr = 0;
			use_regs = 0;
			def_cpsr = 0;
			def_regs = 0;
		end
		`DECODE_LDRSTR:
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rn) | (insn_1a[25] /* I */ ? idxbit(rm) : 0) | (insn_1a[20] /* L */ ? 0 : idxbit(rd));
			def_cpsr = 0;
			def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
		end
		`DECODE_LDMSTM:		/* Block data transfer */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rn) | (insn_1a[20] /* L */ ? 0 : insn_1a[15:0]);
			def_cpsr = insn_1a[22];	/* This is a superset of all cases, anyway. */
			def_regs = (insn_1a[21] /* W */ ? idxbit(rn) : 0) | (insn_1a[20] /* L */ ? insn_1a[15:0] : 0);
		end
		`DECODE_BRANCH:	/* Branch */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = 0;
			def_cpsr = 0;
			def_regs = insn_1a[24] /* L */ ? (16'b1 << 14) : 0;
		end
		`DECODE_LDCSTC:	/* Coprocessor data transfer */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = idxbit(rn);
			def_cpsr = 0;
			def_regs = insn_1a[21] /* W */ ? idxbit(rn) : 0;
		end
		`DECODE_CDP:	/* Coprocessor data op */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = 0;
			def_cpsr = 0;
			def_regs = 0;
		end
		`DECODE_MRCMCR:		/* Coprocessor register transfer */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = insn_1a[20] /* L */ ? 0 : idxbit(rd);
			def_cpsr = 0;
			def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
		end
		`DECODE_SWI:	/* SWI */
		begin
			use_cpsr = `COND_MATTERS(cond);
			use_regs = 0;
			def_cpsr = 0;
			def_regs = 0;
		end
		default:				/* X everything else out */
		begin
			use_cpsr = 1'bx;
			use_regs = 16'bxxxxxxxxxxxxxxxx;
			def_cpsr = 1'bx;
			def_regs = 16'bxxxxxxxxxxxxxxxx;
		end
		endcase
	
	/* Condition checking logic */
	reg condition_met_1a;
	always @(*)
		casez(insn_1a[31:28])
		`COND_EQ:	condition_met_1a = cpsr_1a[`CPSR_Z];
		`COND_NE:	condition_met_1a = !cpsr_1a[`CPSR_Z];
		`COND_CS:	condition_met_1a = cpsr_1a[`CPSR_C];
		`COND_CC:	condition_met_1a = !cpsr_1a[`CPSR_C];
		`COND_MI:	condition_met_1a = cpsr_1a[`CPSR_N];
		`COND_PL:	condition_met_1a = !cpsr_1a[`CPSR_N];
		`COND_VS:	condition_met_1a = cpsr_1a[`CPSR_V];
		`COND_VC:	condition_met_1a = !cpsr_1a[`CPSR_V];
		`COND_HI:	condition_met_1a = cpsr_1a[`CPSR_C] && !cpsr_1a[`CPSR_Z];
		`COND_LS:	condition_met_1a = !cpsr_1a[`CPSR_C] || cpsr_1a[`CPSR_Z];
		`COND_GE:	condition_met_1a = cpsr_1a[`CPSR_N] == cpsr_1a[`CPSR_V];
		`COND_LT:	condition_met_1a = cpsr_1a[`CPSR_N] != cpsr_1a[`CPSR_V];
		`COND_GT:	condition_met_1a = !cpsr_1a[`CPSR_Z] && (cpsr_1a[`CPSR_N] == cpsr_1a[`CPSR_V]);
		`COND_LE:	condition_met_1a = cpsr_1a[`CPSR_Z] || (cpsr_1a[`CPSR_N] != cpsr_1a[`CPSR_V]);
		`COND_AL:	condition_met_1a = 1;
		`COND_NV:	condition_met_1a = 0;
		default:	condition_met_1a = 1'bx;
		endcase
	
	/* Issue logic */
	/* Once it's hit writeback, it's hit the regfile via forwarding so you're done. */
	reg        cpsr_inflight_2a = 0, cpsr_inflight_3a = 0;
	reg [15:0] regs_inflight_2a = 0, regs_inflight_3a = 0;
	
	wire waiting_cpsr_1a = use_cpsr & (cpsr_inflight_2a | cpsr_inflight_3a);
	wire waiting_regs_1a = |(use_regs & (regs_inflight_2a | regs_inflight_3a));
	wire waiting_1a = waiting_cpsr_1a | waiting_regs_1a;
	assign stall_0a = (waiting_1a && !bubble_1a && !flush) || stall_1a;

	reg delayedflush = 0;
	always @(posedge clk/* or negedge Nrst*/)
		if (!Nrst)
			delayedflush <= 0;
		else if (flush && stall_0a /* halp! I can't do it now, maybe later? */)
			delayedflush <= 1;
		else if (!stall_0a /* anything has been handled this time around */)
			delayedflush <= 0;

	/* Actually do the issue. */
	always @(posedge clk or negedge Nrst)
	begin
		if (waiting_1a)
			$display("ISSUE: Stalling instruction %08x because %d/%d", insn_1a, waiting_cpsr_1a, waiting_regs_1a);

		if (!Nrst) begin
			bubble_2a <= 1;
			/*AUTORESET*/
			// Beginning of autoreset for uninitialized flops
			cpsr_inflight_2a <= 1'h0;
			cpsr_inflight_3a <= 1'h0;
			insn_2a <= 32'h0;
			pc_2a <= 32'h0;
			regs_inflight_2a <= 16'h0;
			regs_inflight_3a <= 16'h0;
			// End of automatics
		end else if (!stall_1a)
		begin
			cpsr_inflight_3a <= cpsr_inflight_2a;	/* I'm not sure how well selects work with arrays, and that seems like a dumb thing to get anusulated by. */
			cpsr_inflight_2a <= (waiting_1a || bubble_1a || !condition_met_1a) ? 0 : def_cpsr;
			regs_inflight_3a <= regs_inflight_2a;
			regs_inflight_2a <= (waiting_1a || bubble_1a || !condition_met_1a) ? 0 : def_regs;
			
			bubble_2a <= bubble_1a | waiting_1a | !condition_met_1a | flush | delayedflush;
			pc_2a <= pc_1a;
			insn_2a <= insn_1a;
		end
	end
endmodule
Commit	Line	Data
7aa8541b JW	1	`include "ARM_Constants.v"
7aa8541b JW	2
60cc768b JW	3	module Issue(
60cc768b JW	4	input clk,
09e28f01	5	input Nrst, /* XXX not used yet */
60cc768b	6
ae457ae9	7	input stall_1a, /* pipeline control */
5b3daee2	8	input flush, /* XXX not used yet */
60cc768b	9
a7297aa5 JW	10	input bubble_1a, /* stage inputs */
	11	input [31:0] insn_1a,
	12	input [31:0] pc_1a,
ae457ae9	13	input [31:0] cpsr_1a,
60cc768b	14
ae457ae9 JW	15	output wire stall_0a, /* stage outputs */
	16	output reg bubble_2a = 1,
	17	output reg [31:0] pc_2a = 0,
	18	output reg [31:0] insn_2a = 0
6595c4c8	19	/* XXX other? */
60cc768b JW	20	);
60cc768b JW	21
60cc768b JW	22	`ifdef COPY_PASTA_FODDER
60cc768b JW	23	/* from page 2 of ARM7TDMIvE2.pdf */
a7297aa5	24	casex (insn_1a)
6b8a01fa JW	25	`DECODE_ALU_MULT: /* Multiply -- must come before ALU, because it pattern matches a specific case of ALU */
	26	// `DECODE_ALU_MUL_LONG: /* Multiply long */
	27	`DECODE_ALU_MRS: /* MRS (Transfer PSR to register) */
	28	`DECODE_ALU_MSR: /* MSR (Transfer register to PSR) */
	29	`DECODE_ALU_MSR_FLAGS: /* MSR (Transfer register or immediate to PSR, flag bits only) */
	30	`DECODE_ALU_SWP: /* Atomic swap */
	31	`DECODE_ALU_BX: /* Branch */
	32	`DECODE_ALU_HDATA_REG: /* Halfword transfer - register offset */
	33	`DECODE_ALU_HDATA_IMM: /* Halfword transfer - immediate offset */
	34	`DECODE_ALU: /* ALU */
	35	`DECODE_LDRSTR_UNDEFINED: /* Undefined. I hate ARM */
	36	`DECODE_LDRSTR: /* Single data transfer */
	37	`DECODE_LDMSTM: /* Block data transfer */
	38	`DECODE_BRANCH: /* Branch */
	39	`DECODE_LDCSTC: /* Coprocessor data transfer */
	40	`DECODE_CDP: /* Coprocessor data op */
	41	`DECODE_MRCMCR: /* Coprocessor register transfer */
	42	`DECODE_SWI: /* SWI */
	43	default: /* X everything else out */
60cc768b JW	44	endcase
	45	`endif
	46
e216a134 JW	47	/* Flag setting */
	48	reg use_cpsr;
	49	reg [15:0] use_regs;
	50	reg def_cpsr;
	51	reg [15:0] def_regs;
	52
	53	function [15:0] idxbit;
	54	input [3:0] r;
7aa8541b JW	55	if (r == 15)
	56	idxbit = 0;
	57	else
	58	idxbit = (16'b1) << r;
e216a134 JW	59	endfunction
e216a134 JW	60
a7297aa5 JW	61	wire [3:0] rn = insn_1a[19:16];
	62	wire [3:0] rd = insn_1a[15:12];
	63	wire [3:0] rs = insn_1a[11:8];
	64	wire [3:0] rm = insn_1a[3:0];
	65	wire [3:0] cond = insn_1a[31:28];
e216a134	66
a7297aa5 JW	67	wire [3:0] rd_mul = insn_1a[19:16];
	68	wire [3:0] rn_mul = insn_1a[15:12];
	69	wire [3:0] rs_mul = insn_1a[11:8];
e216a134	70
a7297aa5	71	wire [3:0] alu_opc = insn_1a[24:21];
7aa8541b JW	72
	73	function alu_is_logical;
	74	input [3:0] op;
	75
	76	case (op)
	77	`ALU_AND,`ALU_EOR,`ALU_TST,`ALU_TEQ,`ALU_ORR,`ALU_MOV,`ALU_BIC,`ALU_MVN: alu_is_logical = 1;
	78	default: alu_is_logical = 0;
	79	endcase
	80	endfunction
	81
	82	function alu_flags_only;
	83	input [3:0] op;
	84
	85	case (op)
	86	`ALU_TST,`ALU_TEQ,`ALU_CMP,`ALU_CMN: alu_flags_only = 1;
	87	default: alu_flags_only = 0;
	88	endcase
	89	endfunction
	90
	91	function shift_requires_carry;
	92	input [7:0] shift;
	93
	94	case(shift[1:0])
	95	`SHIFT_LSL: shift_requires_carry = (shift[7:2] == 0);
	96	`SHIFT_LSR: shift_requires_carry = 0;
	97	`SHIFT_ASR: shift_requires_carry = 0;
	98	`SHIFT_ROR: shift_requires_carry = (shift[7:2] == 0);
	99	endcase
	100	endfunction
	101
e216a134	102	always @(*)
a7297aa5	103	casez (insn_1a)
6b8a01fa	104	`DECODE_ALU_MULT: /* Multiply -- must come before ALU, because it pattern matches a specific case of ALU */
e216a134 JW	105	begin
e216a134 JW	106	use_cpsr = `COND_MATTERS(cond);
a7297aa5 JW	107	use_regs = (insn_1a[21] /* accum */ ? idxbit(rn_mul) : 0) \| idxbit(rs_mul) \| idxbit(rm);
a7297aa5 JW	108	def_cpsr = insn_1a[20] /* setcc */;
e216a134 JW	109	def_regs = idxbit(rd_mul);
e216a134 JW	110	end
6b8a01fa JW	111	// `DECODE_ALU_MUL_LONG: /* Multiply long */
6b8a01fa JW	112	`DECODE_ALU_MRS: /* MRS (Transfer PSR to register) */
7aa8541b	113	begin
a7297aa5	114	use_cpsr = `COND_MATTERS(cond) \|\| (insn_1a[22] == 0) /* Source = CPSR */;
7aa8541b JW	115	use_regs = 0;
	116	def_cpsr = 0;
	117	def_regs = idxbit(rd);
	118	end
6b8a01fa	119	`DECODE_ALU_MSR: /* MSR (Transfer register to PSR) */
7aa8541b JW	120	begin
	121	use_cpsr = `COND_MATTERS(cond);
	122	use_regs = idxbit(rm);
	123	def_cpsr = 1;
	124	def_regs = 0;
	125	end
6b8a01fa	126	`DECODE_ALU_MSR_FLAGS: /* MSR (Transfer register or immediate to PSR, flag bits only) */
7aa8541b JW	127	begin
7aa8541b JW	128	use_cpsr = `COND_MATTERS(cond);
a7297aa5	129	use_regs = insn_1a[25] ? 0 : idxbit(rm);
7aa8541b JW	130	def_cpsr = 1;
	131	def_regs = 0;
	132	end
6b8a01fa	133	`DECODE_ALU_SWP: /* Atomic swap */
7aa8541b JW	134	begin
	135	use_cpsr = `COND_MATTERS(cond);
	136	use_regs = idxbit(rn) \| idxbit(rm);
	137	def_cpsr = 0;
	138	def_regs = idxbit(rd);
	139	end
6b8a01fa	140	`DECODE_ALU_BX: /* Branch */
7aa8541b JW	141	begin
	142	use_cpsr = `COND_MATTERS(cond);
	143	use_regs = idxbit(rm);
	144	def_cpsr = 0; // don't care, we'll never get there
	145	def_regs = 0;
	146	end
6b8a01fa	147	`DECODE_ALU_HDATA_REG: /* Halfword transfer - register offset */
7aa8541b JW	148	begin
7aa8541b JW	149	use_cpsr = `COND_MATTERS(cond);
a7297aa5	150	use_regs = idxbit(rn) \| idxbit(rm) \| (insn_1a[20] /* L */ ? 0 : idxbit(rd));
7aa8541b	151	def_cpsr = 0;
a7297aa5	152	def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
7aa8541b	153	end
6b8a01fa	154	`DECODE_ALU_HDATA_IMM: /* Halfword transfer - immediate offset */
7aa8541b JW	155	begin
7aa8541b JW	156	use_cpsr = `COND_MATTERS(cond);
a7297aa5	157	use_regs = idxbit(rn) \| (insn_1a[20] /* L */ ? 0 : idxbit(rd));
7aa8541b	158	def_cpsr = 0;
a7297aa5	159	def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
7aa8541b	160	end
0ede28ef JW	161	`DECODE_ALU: /* ALU */
0ede28ef JW	162	begin
a7297aa5	163	use_cpsr = `COND_MATTERS(cond) \| (!insn_1a[25] /* I */ && shift_requires_carry(insn_1a[11:4]));
0ede28ef	164	use_regs =
a7297aa5 JW	165	(insn_1a[25] /* I */ ? 0 :
a7297aa5 JW	166	(insn_1a[4] /* shift by reg */ ?
0ede28ef JW	167	(idxbit(rs) \| idxbit(rm)) :
	168	(idxbit(rm)))) \|
	169	(((alu_opc != `ALU_MOV) && (alu_opc != `ALU_MVN)) ? idxbit(rn) : 0);
a7297aa5	170	def_cpsr = insn_1a[20] /* S */;
0ede28ef JW	171	def_regs = alu_flags_only(alu_opc) ? 0 : idxbit(rd);
0ede28ef JW	172	end
6b8a01fa	173	`DECODE_LDRSTR_UNDEFINED: /* Undefined. I hate ARM */
7aa8541b JW	174	begin
	175	use_cpsr = 0;
	176	use_regs = 0;
	177	def_cpsr = 0;
	178	def_regs = 0;
	179	end
6b8a01fa JW	180	`DECODE_LDRSTR:
	181	begin
	182	use_cpsr = `COND_MATTERS(cond);
a7297aa5	183	use_regs = idxbit(rn) \| (insn_1a[25] /* I / ? idxbit(rm) : 0) \| (insn_1a[20] / L */ ? 0 : idxbit(rd));
6b8a01fa	184	def_cpsr = 0;
a7297aa5	185	def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
6b8a01fa JW	186	end
6b8a01fa JW	187	`DECODE_LDMSTM: /* Block data transfer */
7aa8541b JW	188	begin
7aa8541b JW	189	use_cpsr = `COND_MATTERS(cond);
a7297aa5 JW	190	use_regs = idxbit(rn) \| (insn_1a[20] /* L */ ? 0 : insn_1a[15:0]);
	191	def_cpsr = insn_1a[22]; /* This is a superset of all cases, anyway. */
	192	def_regs = (insn_1a[21] /* W / ? idxbit(rn) : 0) \| (insn_1a[20] / L */ ? insn_1a[15:0] : 0);
7aa8541b	193	end
6b8a01fa	194	`DECODE_BRANCH: /* Branch */
7aa8541b JW	195	begin
	196	use_cpsr = `COND_MATTERS(cond);
	197	use_regs = 0;
	198	def_cpsr = 0;
a7297aa5	199	def_regs = insn_1a[24] /* L */ ? (16'b1 << 14) : 0;
7aa8541b	200	end
6b8a01fa	201	`DECODE_LDCSTC: /* Coprocessor data transfer */
7aa8541b JW	202	begin
	203	use_cpsr = `COND_MATTERS(cond);
	204	use_regs = idxbit(rn);
	205	def_cpsr = 0;
a7297aa5	206	def_regs = insn_1a[21] /* W */ ? idxbit(rn) : 0;
7aa8541b	207	end
6b8a01fa	208	`DECODE_CDP: /* Coprocessor data op */
7aa8541b JW	209	begin
	210	use_cpsr = `COND_MATTERS(cond);
	211	use_regs = 0;
	212	def_cpsr = 0;
	213	def_regs = 0;
	214	end
6b8a01fa	215	`DECODE_MRCMCR: /* Coprocessor register transfer */
7aa8541b JW	216	begin
7aa8541b JW	217	use_cpsr = `COND_MATTERS(cond);
a7297aa5	218	use_regs = insn_1a[20] /* L */ ? 0 : idxbit(rd);
7aa8541b	219	def_cpsr = 0;
a7297aa5	220	def_regs = insn_1a[20] /* L */ ? idxbit(rd) : 0;
7aa8541b	221	end
6b8a01fa	222	`DECODE_SWI: /* SWI */
7aa8541b JW	223	begin
	224	use_cpsr = `COND_MATTERS(cond);
	225	use_regs = 0;
	226	def_cpsr = 0;
	227	def_regs = 0;
	228	end
e216a134	229	default: /* X everything else out */
7aa8541b JW	230	begin
	231	use_cpsr = 1'bx;
	232	use_regs = 16'bxxxxxxxxxxxxxxxx;
	233	def_cpsr = 1'bx;
	234	def_regs = 16'bxxxxxxxxxxxxxxxx;
	235	end
e216a134	236	endcase
ae281afd JW	237
ae281afd JW	238	/* Condition checking logic */
ae457ae9	239	reg condition_met_1a;
ae281afd	240	always @(*)
a7297aa5	241	casez(insn_1a[31:28])
ae457ae9 JW	242	`COND_EQ: condition_met_1a = cpsr_1a[`CPSR_Z];
	243	`COND_NE: condition_met_1a = !cpsr_1a[`CPSR_Z];
	244	`COND_CS: condition_met_1a = cpsr_1a[`CPSR_C];
	245	`COND_CC: condition_met_1a = !cpsr_1a[`CPSR_C];
	246	`COND_MI: condition_met_1a = cpsr_1a[`CPSR_N];
	247	`COND_PL: condition_met_1a = !cpsr_1a[`CPSR_N];
	248	`COND_VS: condition_met_1a = cpsr_1a[`CPSR_V];
	249	`COND_VC: condition_met_1a = !cpsr_1a[`CPSR_V];
	250	`COND_HI: condition_met_1a = cpsr_1a[`CPSR_C] && !cpsr_1a[`CPSR_Z];
	251	`COND_LS: condition_met_1a = !cpsr_1a[`CPSR_C] \|\| cpsr_1a[`CPSR_Z];
	252	`COND_GE: condition_met_1a = cpsr_1a[`CPSR_N] == cpsr_1a[`CPSR_V];
	253	`COND_LT: condition_met_1a = cpsr_1a[`CPSR_N] != cpsr_1a[`CPSR_V];
	254	`COND_GT: condition_met_1a = !cpsr_1a[`CPSR_Z] && (cpsr_1a[`CPSR_N] == cpsr_1a[`CPSR_V]);
	255	`COND_LE: condition_met_1a = cpsr_1a[`CPSR_Z] \|\| (cpsr_1a[`CPSR_N] != cpsr_1a[`CPSR_V]);
	256	`COND_AL: condition_met_1a = 1;
	257	`COND_NV: condition_met_1a = 0;
	258	default: condition_met_1a = 1'bx;
ae281afd JW	259	endcase
	260
	261	/* Issue logic */
ae457ae9 JW	262	/* Once it's hit writeback, it's hit the regfile via forwarding so you're done. */
	263	reg cpsr_inflight_2a = 0, cpsr_inflight_3a = 0;
	264	reg [15:0] regs_inflight_2a = 0, regs_inflight_3a = 0;
6595c4c8	265
ae457ae9 JW	266	wire waiting_cpsr_1a = use_cpsr & (cpsr_inflight_2a \| cpsr_inflight_3a);
	267	wire waiting_regs_1a = \|(use_regs & (regs_inflight_2a \| regs_inflight_3a));
	268	wire waiting_1a = waiting_cpsr_1a \| waiting_regs_1a;
	269	assign stall_0a = (waiting_1a && !bubble_1a && !flush) \|\| stall_1a;
c4e2ac3b JW	270
c4e2ac3b JW	271	reg delayedflush = 0;
a4f724e6 JW	272	always @(posedge clk/* or negedge Nrst*/)
	273	if (!Nrst)
	274	delayedflush <= 0;
ae457ae9	275	else if (flush && stall_0a /* halp! I can't do it now, maybe later? */)
c4e2ac3b	276	delayedflush <= 1;
ae457ae9	277	else if (!stall_0a /* anything has been handled this time around */)
c4e2ac3b	278	delayedflush <= 0;
a02d953f	279
6595c4c8	280	/* Actually do the issue. */
a4f724e6	281	always @(posedge clk or negedge Nrst)
6595c4c8	282	begin
ae457ae9 JW	283	if (waiting_1a)
ae457ae9 JW	284	$display("ISSUE: Stalling instruction %08x because %d/%d", insn_1a, waiting_cpsr_1a, waiting_regs_1a);
149bcd1a	285
a4f724e6	286	if (!Nrst) begin
ae457ae9	287	bubble_2a <= 1;
58b94135 JW	288	/AUTORESET/
	289	// Beginning of autoreset for uninitialized flops
	290	cpsr_inflight_2a <= 1'h0;
	291	cpsr_inflight_3a <= 1'h0;
	292	insn_2a <= 32'h0;
	293	pc_2a <= 32'h0;
	294	regs_inflight_2a <= 16'h0;
	295	regs_inflight_3a <= 16'h0;
	296	// End of automatics
ae457ae9	297	end else if (!stall_1a)
4ddc4b72	298	begin
ae457ae9 JW	299	cpsr_inflight_3a <= cpsr_inflight_2a; /* I'm not sure how well selects work with arrays, and that seems like a dumb thing to get anusulated by. */
	300	cpsr_inflight_2a <= (waiting_1a \|\| bubble_1a \|\| !condition_met_1a) ? 0 : def_cpsr;
	301	regs_inflight_3a <= regs_inflight_2a;
	302	regs_inflight_2a <= (waiting_1a \|\| bubble_1a \|\| !condition_met_1a) ? 0 : def_regs;
4ddc4b72	303
ae457ae9 JW	304	bubble_2a <= bubble_1a \| waiting_1a \| !condition_met_1a \| flush \| delayedflush;
	305	pc_2a <= pc_1a;
	306	insn_2a <= insn_1a;
1c2e57dc	307	end
6595c4c8	308	end
60cc768b	309	endmodule