Context Navigation

os2wb.v @ 36

Revision 36, 60.6 KB checked in by pntsvt00, 14 years ago (diff)
modificati i file per la compilazione. boot.s linkato a 0xFFF0000020

Rev	Line
[10]	1	`timescale 1ns / 1ps
	2	//////////////////////////////////////////////////////////////////////////////////
	3	// Company: (C) Athree, 2009
	4	// Engineer: Dmitry Rozhdestvenskiy
	5	// Email [email protected] [email protected] [email protected]
	6	//
	7	// Design Name: Bridge from SPARC Core to Wishbone Master
	8	// Module Name: os2wb
	9	// Project Name: SPARC SoC single-core
	10	//
	11	// LICENSE:
	12	// This is a Free Hardware Design; you can redistribute it and/or
	13	// modify it under the terms of the GNU General Public License
	14	// version 2 as published by the Free Software Foundation.
	15	// The above named program is distributed in the hope that it will
	16	// be useful, but WITHOUT ANY WARRANTY; without even the implied
	17	// warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
	18	// See the GNU General Public License for more details.
	19	//
	20	//////////////////////////////////////////////////////////////////////////////////
	21	module os2wb(
	22	input clk,
	23	input rstn,
	24
	25	// Core interface
[26]	26	input [ 4:0] pcx_req,
[10]	27	input pcx_atom,
[26]	28	input [123:0] pcx_data,
	29
[10]	30	output reg [ 4:0] pcx_grant,
	31	output reg cpx_ready,
	32	output reg [144:0] cpx_packet,
	33
	34	// Wishbone master interface
	35	input [ 63:0] wb_data_i,
	36	input wb_ack,
	37	output reg wb_cycle,
	38	output reg wb_strobe,
	39	output reg wb_we,
	40	output reg [ 7:0] wb_sel,
	41	output reg [ 63:0] wb_addr,
	42	output reg [ 63:0] wb_data_o,
	43
	44	// FPU interface
	45	output reg [123:0] fp_pcx,
	46	output reg fp_req,
	47	input [144:0] fp_cpx,
	48	input fp_rdy,
	49
	50	// Ethernet interrupt, sensed on posedge, mapped to vector 'd29
	51	input eth_int
	52	);
	53
	54	reg [123:0] pcx_packet_d; // Latched incoming PCX packet
	55	reg [123:0] pcx_packet_2nd; // Second packet for atomic (CAS)
	56	reg [ 4:0] pcx_req_d; // Latched request
	57	reg pcx_atom_d; // Latched atomic flasg
	58	reg [ 4:0] state; // FSM state
	59	reg [144:0] cpx_packet_1; // First CPX packet
	60	reg [144:0] cpx_packet_2; // Second CPX packet (for atomics and cached IFILLs)
	61	reg cpx_two_packet; // CPX answer is two-packet (!=atomic, SWAP has atomic==0 and answer is two-packet)
	62
	63	reg [ 3:0] inval_vect0; // Invalidate, instr/data, way
	64	reg [ 3:0] inval_vect1; // IFill may cause two D lines invalidation at a time
	65
	66	wire [111:0] store_inv_vec; // Store invalidation vector
	67
	68	assign store_inv_vec[111:91]=0;
	69	assign store_inv_vec[90:88]=((pcx_packet_d[64+5:64+4]==2'b11) && inval_vect0[3:2]==2'b11) ? {inval_vect0[1:0],1'b1}:3'b000;
	70	assign store_inv_vec[87:60]=0;
	71	assign store_inv_vec[59:56]=((pcx_packet_d[64+5:64+4]==2'b10) && inval_vect0[3:2]==2'b11) \|\| ((pcx_packet_d[64+5]==1'b1) && inval_vect0[3:2]==2'b10) ? {inval_vect0[1:0],!inval_vect0[2],inval_vect0[2]}:4'b0000;
	72	assign store_inv_vec[55:35]=0;
	73	assign store_inv_vec[34:32]=((pcx_packet_d[64+5:64+4]==2'b01) && inval_vect0[3:2]==2'b11) ? {inval_vect0[1:0],1'b1}:3'b000;
	74	assign store_inv_vec[31:4]=0;
	75	assign store_inv_vec[3:0]=((pcx_packet_d[64+5:64+4]==2'b00) && inval_vect0[3:2]==2'b11) \|\| ((pcx_packet_d[64+5]==1'b0) && inval_vect0[3:2]==2'b10) ? {inval_vect0[1:0],!inval_vect0[2],inval_vect0[2]}:4'b0000;
	76
	77	wire [28:0] dcache0_do0;
	78	wire [28:0] dcache0_do1;
	79	wire [28:0] dcache1_do0;
	80	wire [28:0] dcache1_do1;
	81	wire [28:0] dcache2_do0;
	82	wire [28:0] dcache2_do1;
	83	wire [28:0] dcache3_do0;
	84	wire [28:0] dcache3_do1;
	85	wire [28:0] icache0_do;
	86	wire [28:0] icache1_do;
	87	wire [28:0] icache2_do;
	88	wire [28:0] icache3_do;
	89
	90	`define TEST_DRAM_1 5'b00000
	91	`define TEST_DRAM_2 5'b00001
	92	`define TEST_DRAM_3 5'b00010
	93	`define TEST_DRAM_4 5'b00011
	94	`define INIT_DRAM_1 5'b00100
	95	`define INIT_DRAM_2 5'b00101
	96	`define WAKEUP 5'b00110
	97	`define PCX_IDLE 5'b00111
	98	`define GOT_PCX_REQ 5'b01000
	99	`define PCX_REQ_2ND 5'b01001
	100	`define PCX_REQ_STEP1 5'b01010
	101	`define PCX_REQ_STEP1_1 5'b01011
	102	`define PCX_REQ_STEP2 5'b01100
	103	`define PCX_REQ_STEP2_1 5'b01101
	104	`define PCX_REQ_STEP3 5'b01110
	105	`define PCX_REQ_STEP3_1 5'b01111
	106	`define PCX_REQ_STEP4 5'b10000
	107	`define PCX_REQ_STEP4_1 5'b10001
	108	`define PCX_BIS 5'b10010
	109	`define PCX_BIS_1 5'b10011
	110	`define PCX_BIS_2 5'b10100
	111	`define CPX_READY_1 5'b10101
	112	`define CPX_READY_2 5'b10110
	113	`define PCX_UNKNOWN 5'b11000
	114	`define PCX_FP_1 5'b11001
	115	`define PCX_FP_2 5'b11010
	116	`define FP_WAIT 5'b11011
	117	`define CPX_FP 5'b11100
	118	`define CPX_SEND_ETH_IRQ 5'b11101
	119	`define CPX_INT_VEC_DIS 5'b11110
	120	`define PCX_REQ_CAS_COMPARE 5'b11111
	121
[22]	122	//`define MEM_SIZE 64'h00000000_10000000 //256 MB
	123	//`define MEM_SIZE 64'h00000000_00100000 //1MB
[33]	124	`define MEM_SIZE 64'h00000000_00001000 //16KB
[10]	125
[22]	126	// sal: escludo test della DRAM `define TEST_DRAM 1
[27]	127	`define TEST_DRAM 0
[10]	128	`define DEBUGGING 1
	129
	130	reg cache_init;
	131	wire [3:0] dcache0_hit;
	132	wire [3:0] dcache1_hit;
	133	wire [3:0] icache_hit;
	134	reg multi_hit;
	135	reg multi_hit1;
	136	reg eth_int_d;
	137	reg eth_int_send;
	138	reg eth_int_sent;
	139	reg [3:0] cnt;
	140
	141	// PCX channel FIFO
	142	wire [129:0] pcx_data_fifo;
	143	wire pcx_fifo_empty;
	144	reg [ 4:0] pcx_req_1;
	145	reg [ 4:0] pcx_req_2;
	146	reg pcx_atom_1;
	147	reg pcx_atom_2;
	148	reg pcx_data_123_d;
	149
	150	always @(posedge clk)
	151	begin
	152	pcx_req_1<=pcx_req;
	153	pcx_atom_1<=pcx_atom;
	154	pcx_atom_2<=pcx_atom_1;
	155	pcx_req_2<=pcx_atom_1 ? pcx_req_1:5'b0;
	156	pcx_grant<=(pcx_req_1 \| pcx_req_2);
	157	pcx_data_123_d<=pcx_data[123];
	158	end
	159
	160	/*pcx_fifo pcx_fifo_inst(
	161	// FIFO should be first word fall-through
	162	// It has no full flag as the core will send only limited number of requests,
	163	// in original design we used it 32 words deep
	164	// Just make it deeper if you experience overflow -
	165	// you can't just send no grant on full because the core expects immediate
	166	// grant for at least two requests for each zone
	167	.aclr(!rstn),
	168	.clock(clk),
	169	.data({pcx_atom_1,pcx_req_1,pcx_data}),
	170	.rdreq(fifo_rd),
	171	.wrreq((pcx_req_1!=5'b00000 && pcx_data[123]) \|\| (pcx_atom_2 && pcx_data_123_d)),
	172	// Second atomic packet for FPU may be invalid, but should be sent to FPU
	173	// so if the first atomic packet is valid we latch both
	174	.empty(pcx_fifo_empty),
	175	.q(pcx_data_fifo)
	176	);
	177	*/
[14]	178	reg fifo_rd;
	179	wire [123:0] pcx_packet;
	180
	181
[10]	182	pcx_fifo pcx_fifo_inst(
	183	.clk(clk),
[22]	184	.rst(!rstn),
[10]	185	.din({pcx_atom_1,pcx_req_1,pcx_data}),
	186	.rd_en(fifo_rd),
	187	.wr_en((pcx_req_1!=5'b00000 && pcx_data[123]) \|\| (pcx_atom_2 && pcx_data_123_d)),
	188	.empty(pcx_fifo_empty),
[22]	189	.full(),
[10]	190	.dout(pcx_data_fifo)
	191	);
	192
	193
	194
	195	// --------------------------
	196
	197	always @(posedge clk or negedge rstn)
	198	if(!rstn)
	199	eth_int_send<=0;
	200	else
	201	begin
	202	eth_int_d<=eth_int;
	203	if(eth_int && !eth_int_d)
	204	eth_int_send<=1;
	205	else
	206	if(eth_int_sent)
	207	eth_int_send<=0;
	208	end
	209
[14]	210
[10]	211	assign pcx_packet=pcx_data_fifo[123:0];
	212
	213	always @(posedge clk or negedge rstn)
	214	if(rstn==0)
	215	begin
[22]	216	//$display("INFO: OS2WB: RST_DRAM at %t",$time);
[10]	217	if(`TEST_DRAM)
	218	state<=`TEST_DRAM_1;
	219	else
[22]	220	state<=`INIT_DRAM_1; // DRAM initialization is mandatory!
[10]	221	cpx_ready<=0;
	222	fifo_rd<=0;
	223	cpx_packet<=145'b0;
	224	wb_cycle<=0;
	225	wb_strobe<=0;
	226	wb_we<=0;
	227	wb_sel<=0;
	228	wb_addr<=64'b0;
	229	wb_data_o<=64'b0;
	230	pcx_packet_d<=124'b0;
	231	fp_pcx<=124'b0;
	232	fp_req<=0;
	233	end
	234	else
	235	case(state)
	236	`TEST_DRAM_1:
	237	begin
[26]	238	//$display("INFO: OS2WB: TEST_DRAM_1");
[10]	239	wb_cycle<=1;
	240	wb_strobe<=1;
	241	wb_sel<=8'hFF;
	242	wb_we<=1;
	243	state<=`TEST_DRAM_2;
	244	end
	245	`TEST_DRAM_2:
	246	if(wb_ack)
	247	begin
[26]	248	//$display("INFO: OS2WB: TEST_DRAM_2 at time %d with wb_addr=%x",$time,wb_addr);
[10]	249	wb_strobe<=0;
	250	if(wb_addr<`MEM_SIZE-8)
	251	begin
	252	wb_addr[31:0]<=wb_addr[31:0]+8;
	253	wb_data_o<={wb_addr[31:0]+8,wb_addr[31:0]+8};
	254	state<=`TEST_DRAM_1;
	255	end
	256	else
	257	begin
	258	state<=`TEST_DRAM_3;
	259	wb_cycle<=0;
	260	wb_sel<=0;
	261	wb_we<=0;
	262	wb_data_o<=64'b0;
	263	wb_addr<=64'b0;
	264	end
	265	end
	266	`TEST_DRAM_3:
	267	begin
[26]	268	//$display("INFO: OS2WB: TEST_DRAM_3");
[10]	269	wb_cycle<=1;
	270	wb_strobe<=1;
	271	wb_sel<=8'hFF;
	272	state<=`TEST_DRAM_4;
	273	end
	274	`TEST_DRAM_4:
	275	if(wb_ack)
	276	begin
[26]	277	// $display("INFO: OS2WB: TEST_DRAM_4 at %t",$time);
[10]	278	wb_strobe<=0;
	279	if(wb_addr<`MEM_SIZE-8)
	280	begin
[23]	281	//if(wb_data_i=={wb_addr[31:0],wb_addr[31:0]})
	282	// begin
[10]	283	wb_addr[31:0]<=wb_addr[31:0]+8;
	284	state<=`TEST_DRAM_3;
[23]	285	// end
	286	//else
[26]	287	// $display("expected %x, obtained %x",{wb_addr[31:0],wb_addr[31:0]},wb_data_i);
[10]	288	end
	289	else
	290	begin
[26]	291	//$display("INFO: OS2WB: INIT_DRAM at %t",$time);
[10]	292	state<=`INIT_DRAM_1;
	293	wb_cycle<=0;
	294	wb_sel<=0;
	295	wb_we<=0;
	296	wb_data_o<=64'b0;
	297	wb_addr<=64'b0;
	298	end
	299	end
	300	`INIT_DRAM_1:
	301	begin
	302	wb_cycle<=1;
	303	wb_strobe<=1;
	304	wb_sel<=8'hFF;
	305	wb_we<=1;
	306	cache_init<=1; // We also init cache directories here
	307	state<=`INIT_DRAM_2;
	308	end
	309	`INIT_DRAM_2:
	310	if(wb_ack)
	311	begin
	312	wb_strobe<=0;
	313	if(wb_addr<`MEM_SIZE-8)
	314	begin
[22]	315	//for debug
	316	// if (wb_addr[10:3]==8'b0)
	317	// $display("INFO: OS2WB: INIT_DRAM_2 at time %d with wb_addr=%x",$time,wb_addr);
	318	//
[10]	319	wb_addr[31:0]<=wb_addr[31:0]+8;
	320	pcx_packet_d[64+11:64+4]<=pcx_packet_d[64+11:64+4]+1; // Address for cachedir init
	321	state<=`INIT_DRAM_1;
	322	end
	323	else
	324	begin
[26]	325	//$display("INFO: OS2WB: WAKEUP_DRAM at %t",$time);
[10]	326	state<=`WAKEUP;
	327	wb_cycle<=0;
	328	wb_sel<=0;
	329	wb_we<=0;
	330	cache_init<=0;
	331	wb_addr<=64'b0;
	332	end
	333	end
	334	`WAKEUP:
	335	begin
	336	cpx_packet<=145'h1700000000000000000000000000000010001;
	337	cpx_ready<=1;
	338	state<=`PCX_IDLE;
	339	end
	340	`PCX_IDLE:
	341	begin
	342	cnt<=0;
	343	cpx_packet<=145'b0;
	344	cpx_ready<=0;
	345	cpx_two_packet<=0;
[27]	346	//inval_vect0[3]<=0;
	347	//inval_vect1[3]<=0;
	348	inval_vect0<=0;
	349	inval_vect1<=0;
[10]	350	multi_hit<=0;
	351	multi_hit1<=0;
	352	if(eth_int_send)
	353	begin
	354	state<=`CPX_SEND_ETH_IRQ;
	355	eth_int_sent<=1;
	356	end
	357	else
	358	if(!pcx_fifo_empty)
	359	begin
	360	pcx_req_d<=pcx_data_fifo[128:124];
	361	pcx_atom_d<=pcx_data_fifo[129];
	362	fifo_rd<=1;
	363	state<=`GOT_PCX_REQ;
	364	end
	365	end
	366	`GOT_PCX_REQ:
	367	begin
	368	pcx_packet_d<=pcx_packet;
	369	if(`DEBUGGING)
	370	begin
[22]	371	$display("INFO: OS2WB: GOT_PCX_REQ");
[10]	372	wb_sel[1:0]<=pcx_packet[113:112];
	373	wb_sel[2]<=1;
	374	end
	375	if(pcx_packet[103:64]==40'h9800000800 && pcx_packet[122:118]==5'b00001)
	376	begin
	377	state<=`CPX_INT_VEC_DIS;
	378	fifo_rd<=0;
	379	end
	380	else
	381	if(pcx_atom_d==0)
	382	begin
	383	fifo_rd<=0;
	384	if(pcx_packet[122:118]==5'b01010) // FP req
	385	begin
	386	state<=`PCX_FP_1;
	387	pcx_packet_2nd[123]<=0;
	388	end
	389	else
	390	state<=`PCX_REQ_STEP1;
	391	end
	392	else
	393	state<=`PCX_REQ_2ND;
	394	end
	395	`PCX_REQ_2ND:
	396	begin
[26]	397	$display("INFO: OS2WB: GOT_PCX_REQ_2ND");
[10]	398	pcx_packet_2nd<=pcx_packet; //Latch second packet for atomics
	399	if(`DEBUGGING)
	400	if(pcx_fifo_empty)
	401	wb_sel<=8'h67;
	402	fifo_rd<=0;
	403	if(pcx_packet_d[122:118]==5'b01010) // FP req
	404	state<=`PCX_FP_1;
	405	else
	406	state<=`PCX_REQ_STEP1;
	407	end
	408	`PCX_REQ_STEP1:
	409	begin
	410	if(pcx_packet_d[111]==1'b1) // Invalidate request
	411	begin
[26]	412	$display("INFO: OS2WB: INVALIDATE");
[10]	413	cpx_packet_1[144]<=1; // Valid
	414	cpx_packet_1[143:140]<=4'b0100; // Invalidate reply is Store ACK
	415	cpx_packet_1[139]<=1; // L2 miss
	416	cpx_packet_1[138:137]<=0; // Error
	417	cpx_packet_1[136]<=pcx_packet_d[117]; // Non-cacheble
	418	cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID
	419	cpx_packet_1[133:131]<=0; // Way valid
	420	cpx_packet_1[130]<=((pcx_packet_d[122:118]==5'b10000) && (pcx_req_d==5'b10000)) ? 1:0; // Four byte fill
	421	cpx_packet_1[129]<=pcx_atom_d;
	422	cpx_packet_1[128]<=pcx_packet_d[110]; // Prefetch
	423	cpx_packet_1[127:0]<={2'b0,pcx_packet_d[109]/BIS/,pcx_packet_d[122:118]==5'b00000 ? 2'b01:2'b10,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],112'b0};
	424	state<=`CPX_READY_1;
	425	end
	426	else
	427	if(pcx_packet_d[122:118]!=5'b01001) // Not INT
	428	begin
[26]	429	$display("INFO: OS2WB: PCX_REQ_STEP1");
[10]	430	wb_cycle<=1'b1;
	431	wb_strobe<=1'b1;
	432	if((pcx_packet_d[122:118]==5'b00000 && !pcx_req_d[4]) \|\| pcx_packet_d[122:118]==5'b00010 \|\| pcx_packet_d[122:118]==5'b00100 \|\| pcx_packet_d[122:118]==5'b00110)
[26]	433	begin
	434	$display("INFO: OS2WB: load/streadload ecc");
[10]	435	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+4],4'b0000}; //DRAM load/streamload, CAS and SWAP always use DRAM and load first
[26]	436	end
	437	else
[10]	438	if(pcx_packet_d[122:118]==5'b10000 && !pcx_req_d[4])
[26]	439	begin
[36]	440	$display("INFO: OS2WB: DRAM ifill");
[10]	441	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b00000}; //DRAM ifill
[26]	442	end
	443	else
[10]	444	if(pcx_packet_d[64+39:64+28]==12'hFFF && pcx_packet_d[64+27:64+24]!=4'b0) // flash remap FFF1->FFF8
[36]	445	begin
	446	$display("INFO: OS2WB: flash remap");
	447	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3]+37'h0000E00000,3'b000};
	448	end
	449	else
[10]	450	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3],3'b000};
	451	wb_data_o<=pcx_packet_d[63:0];
	452	state<=`PCX_REQ_STEP1_1;
	453	end
	454	else
	455	if((pcx_packet_d[12:10]!=3'b000) && !pcx_packet_d[117]) // Not FLUSH int and not this core
	456	state<=`PCX_IDLE;
	457	else
	458	state<=`CPX_READY_1;
	459	case(pcx_packet_d[122:118]) // Packet type
	460	5'b00000://Load
	461	begin
[26]	462	$display("INFO: OS2WB: PCX_REQ_STEP1, Load");
[10]	463	wb_we<=0;
	464	if(!pcx_packet_d[110] && !pcx_packet_d[117])
	465	case(icache_hit)
	466	4'b0000:;
	467	4'b0001:inval_vect0<=4'b1_0_00;
	468	4'b0010:inval_vect0<=4'b1_0_01;
	469	4'b0100:inval_vect0<=4'b1_0_10;
	470	4'b1000:inval_vect0<=4'b1_0_11;
	471	default:multi_hit<=1;
	472	endcase
	473	if(!pcx_req_d[4])
	474	wb_sel<=8'b11111111; // DRAM requests are always 128 bit
	475	else
	476	case(pcx_packet_d[106:104]) //Size
	477	3'b000://Byte
	478	case(pcx_packet_d[64+2:64])
	479	3'b000:wb_sel<=8'b10000000;
	480	3'b001:wb_sel<=8'b01000000;
	481	3'b010:wb_sel<=8'b00100000;
	482	3'b011:wb_sel<=8'b00010000;
	483	3'b100:wb_sel<=8'b00001000;
	484	3'b101:wb_sel<=8'b00000100;
	485	3'b110:wb_sel<=8'b00000010;
	486	3'b111:wb_sel<=8'b00000001;
	487	endcase
	488	3'b001://Halfword
	489	case(pcx_packet_d[64+2:64+1])
	490	2'b00:wb_sel<=8'b11000000;
	491	2'b01:wb_sel<=8'b00110000;
	492	2'b10:wb_sel<=8'b00001100;
	493	2'b11:wb_sel<=8'b00000011;
	494	endcase
	495	3'b010://Word
	496	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	497	3'b011://Doubleword
	498	wb_sel<=8'b11111111;
	499	3'b100://Quadword
	500	wb_sel<=8'b11111111;
	501	3'b111://Cacheline
	502	wb_sel<=8'b11111111;
	503	default:
	504	wb_sel<=8'b01011010; // Unreal eye-catching value for debug
	505	endcase
	506	end
	507	5'b00001://Store
	508	begin
[26]	509	$display("INFO: OS2WB: PCX_REQ_STEP1, Store");
[10]	510	wb_we<=1;
	511	case({icache_hit,dcache0_hit})
	512	8'b00000000:;
	513	8'b00000001:inval_vect0<=4'b1_1_00;
	514	8'b00000010:inval_vect0<=4'b1_1_01;
	515	8'b00000100:inval_vect0<=4'b1_1_10;
	516	8'b00001000:inval_vect0<=4'b1_1_11;
	517	8'b00010000:inval_vect0<=4'b1_0_00;
	518	8'b00100000:inval_vect0<=4'b1_0_01;
	519	8'b01000000:inval_vect0<=4'b1_0_10;
	520	8'b10000000:inval_vect0<=4'b1_0_11;
	521	default:multi_hit<=1;
	522	endcase
	523	if(pcx_packet_d[110:109]!=2'b00) //Block (or init) store
	524	wb_sel<=8'b11111111; // Blocks are always 64 bit
	525	else
	526	case(pcx_packet_d[106:104]) //Size
	527	3'b000://Byte
	528	case(pcx_packet_d[64+2:64])
	529	3'b000:wb_sel<=8'b10000000;
	530	3'b001:wb_sel<=8'b01000000;
	531	3'b010:wb_sel<=8'b00100000;
	532	3'b011:wb_sel<=8'b00010000;
	533	3'b100:wb_sel<=8'b00001000;
	534	3'b101:wb_sel<=8'b00000100;
	535	3'b110:wb_sel<=8'b00000010;
	536	3'b111:wb_sel<=8'b00000001;
	537	endcase
	538	3'b001://Halfword
	539	case(pcx_packet_d[64+2:64+1])
	540	2'b00:wb_sel<=8'b11000000;
	541	2'b01:wb_sel<=8'b00110000;
	542	2'b10:wb_sel<=8'b00001100;
	543	2'b11:wb_sel<=8'b00000011;
	544	endcase
	545	3'b010://Word
	546	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	547	3'b011://Doubleword
	548	wb_sel<=8'b11111111;
	549	default:
	550	if(`DEBUGGING)
	551	wb_sel<=8'b01011010; // Unreal eye-catching value for debug
	552	endcase
	553	end
	554	5'b00010://CAS
	555	begin
[26]	556	$display("INFO: OS2WB: PCX_REQ_STEP1, CAS");
[10]	557	wb_we<=0; //Load first
	558	case({icache_hit,dcache0_hit})
	559	8'b00000000:;
	560	8'b00000001:inval_vect0<=4'b1_1_00;
	561	8'b00000010:inval_vect0<=4'b1_1_01;
	562	8'b00000100:inval_vect0<=4'b1_1_10;
	563	8'b00001000:inval_vect0<=4'b1_1_11;
	564	8'b00010000:inval_vect0<=4'b1_0_00;
	565	8'b00100000:inval_vect0<=4'b1_0_01;
	566	8'b01000000:inval_vect0<=4'b1_0_10;
	567	8'b10000000:inval_vect0<=4'b1_0_11;
	568	default:multi_hit<=1;
	569	endcase
	570	wb_sel<=8'b11111111; // CAS loads are as cacheline
	571	end
	572	5'b00100://STRLOAD
	573	begin
[26]	574	$display("INFO: OS2WB: PCX_REQ_STEP1, STRLOAD");
[10]	575	wb_we<=0;
	576	wb_sel<=8'b11111111; // Stream loads are always 128 bit
	577	end
	578	5'b00101://STRSTORE
	579	begin
[26]	580	$display("INFO: OS2WB: PCX_REQ_STEP1, STRSTORE");
[10]	581	wb_we<=1;
	582	case({icache_hit,dcache0_hit})
	583	8'b00000000:;
	584	8'b00000001:inval_vect0<=4'b1_1_00;
	585	8'b00000010:inval_vect0<=4'b1_1_01;
	586	8'b00000100:inval_vect0<=4'b1_1_10;
	587	8'b00001000:inval_vect0<=4'b1_1_11;
	588	8'b00010000:inval_vect0<=4'b1_0_00;
	589	8'b00100000:inval_vect0<=4'b1_0_01;
	590	8'b01000000:inval_vect0<=4'b1_0_10;
	591	8'b10000000:inval_vect0<=4'b1_0_11;
	592	default:multi_hit<=1;
	593	endcase
	594	case(pcx_packet_d[106:104]) //Size
	595	3'b000://Byte
	596	case(pcx_packet_d[64+2:64])
	597	3'b000:wb_sel<=8'b10000000;
	598	3'b001:wb_sel<=8'b01000000;
	599	3'b010:wb_sel<=8'b00100000;
	600	3'b011:wb_sel<=8'b00010000;
	601	3'b100:wb_sel<=8'b00001000;
	602	3'b101:wb_sel<=8'b00000100;
	603	3'b110:wb_sel<=8'b00000010;
	604	3'b111:wb_sel<=8'b00000001;
	605	endcase
	606	3'b001://Halfword
	607	case(pcx_packet_d[64+2:64+1])
	608	2'b00:wb_sel<=8'b11000000;
	609	2'b01:wb_sel<=8'b00110000;
	610	2'b10:wb_sel<=8'b00001100;
	611	2'b11:wb_sel<=8'b00000011;
	612	endcase
	613	3'b010://Word
	614	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	615	3'b011://Doubleword
	616	wb_sel<=8'b11111111;
	617	3'b100://Quadword
	618	wb_sel<=8'b11111111;
	619	3'b111://Cacheline
	620	wb_sel<=8'b11111111;
	621	default:
	622	wb_sel<=8'b01011010; // Unreal eye-catching value for debug
	623	endcase
	624	end
	625	5'b00110://SWAP/LDSTUB
	626	begin
[26]	627	$display("INFO: OS2WB: PCX_REQ_STEP1, SWAP/LDSTUB");
[10]	628	case({icache_hit,dcache0_hit})
	629	8'b00000000:;
	630	8'b00000001:inval_vect0<=4'b1_1_00;
	631	8'b00000010:inval_vect0<=4'b1_1_01;
	632	8'b00000100:inval_vect0<=4'b1_1_10;
	633	8'b00001000:inval_vect0<=4'b1_1_11;
	634	8'b00010000:inval_vect0<=4'b1_0_00;
	635	8'b00100000:inval_vect0<=4'b1_0_01;
	636	8'b01000000:inval_vect0<=4'b1_0_10;
	637	8'b10000000:inval_vect0<=4'b1_0_11;
	638	default:multi_hit<=1;
	639	endcase
	640	wb_we<=0; // Load first, as CAS
	641	wb_sel<=8'b11111111; // SWAP/LDSTUB loads are as cacheline
	642	end
	643	5'b01001://INT
[26]	644	begin
	645	$display("INFO: OS2WB: PCX_REQ_STEP1, INT");
	646	if(pcx_packet_d[117]) // Flush
[10]	647	cpx_packet_1<={9'h171,pcx_packet_d[113:112],11'h0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],30'h0,pcx_packet_d[17:0],46'b0,pcx_packet_d[17:0]}; //FLUSH instruction answer
[26]	648	else // Tread-to-thread interrupt
[10]	649	cpx_packet_1<={9'h170,pcx_packet_d[113:112],52'h0,pcx_packet_d[17:0],46'h0,pcx_packet_d[17:0]};
[26]	650	//5'b01010: FP1 - processed by separate state
	651	//5'b01011: FP2 - processed by separate state
	652	//5'b01101: FWDREQ - not implemented
	653	//5'b01110: FWDREPL - not implemented
	654	end
	655	5'b10000://IFILL
[10]	656	begin
[26]	657	$display("INFO: OS2WB: PCX_REQ_STEP1, IFILL");
[10]	658	wb_we<=0;
	659	if(!pcx_req_d[4]) // not I/O access
	660	begin
	661	case(dcache0_hit)
	662	4'b0000:;
	663	4'b0001:inval_vect0<=4'b1_1_00;
	664	4'b0010:inval_vect0<=4'b1_1_01;
	665	4'b0100:inval_vect0<=4'b1_1_10;
	666	4'b1000:inval_vect0<=4'b1_1_11;
	667	default:multi_hit<=1;
	668	endcase
	669	case(dcache1_hit)
	670	4'b0000:;
	671	4'b0001:inval_vect1<=4'b1_1_00;
	672	4'b0010:inval_vect1<=4'b1_1_01;
	673	4'b0100:inval_vect1<=4'b1_1_10;
	674	4'b1000:inval_vect1<=4'b1_1_11;
	675	default:multi_hit1<=1;
	676	endcase
	677	end
	678	if(pcx_req_d[4]) // I/O access
	679	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	680	else
	681	wb_sel<=8'b11111111;
	682	end
	683	default:
	684	begin
	685	wb_we<=0;
	686	wb_sel<=8'b10101010; // Unreal eye-catching value for debug
	687	end
	688	endcase
	689	end
	690	`PCX_REQ_STEP1_1:
	691	begin
	692	if(wb_ack)
	693	begin
[36]	694	$display("INFO: OS2WB: PCX_REQ_STEP1_1 wb_addr = %x wb_data_i = %x wb_data_o = %x ",wb_addr, wb_data_i,wb_data_o);
[10]	695	cpx_packet_1[144]<=1; // Valid
[26]	696	cpx_packet_1[139]<=(pcx_packet_d[122:118]==5'b00000) \|\| (pcx_packet_d[122:118]==5'b10000) ? 1:0; // L2 always miss on load and ifill
[10]	697	cpx_packet_1[138:137]<=0; // Error
	698	cpx_packet_1[136]<=pcx_packet_d[117] \|\| (pcx_packet_d[122:118]==5'b00001) ? 1:0; // Non-cacheble is set on store too
	699	cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID
	700	if((pcx_packet_d[122:118]==5'b00000 && !pcx_packet_d[117] && !pcx_packet_d[110]) \|\| (pcx_packet_d[122:118]==5'b10000)) // Cacheble Load or IFill
	701	cpx_packet_1[133:131]<={inval_vect0[3],inval_vect0[1:0]};
	702	else
	703	cpx_packet_1[133:131]<=3'b000; // Way valid
	704	if(pcx_packet_d[122:118]==5'b00100) // Strload
	705	cpx_packet_1[130]<=pcx_packet_d[106]; // A
	706	else
	707	if(pcx_packet_d[122:118]==5'b00101) // Stream store
	708	cpx_packet_1[130]<=pcx_packet_d[108]; // A
	709	else
	710	cpx_packet_1[130]<=((pcx_packet_d[122:118]==5'b10000) && pcx_req_d[4]) ? 1:0; // Four byte fill
	711	if(pcx_packet_d[122:118]==5'b00100) // Strload
	712	cpx_packet_1[129]<=pcx_packet_d[105]; // B
	713	else
	714	cpx_packet_1[129]<=pcx_atom_d \|\| (pcx_packet_d[122:118]==5'b00110); // SWAP is single-packet but needs atom in CPX
	715	cpx_packet_1[128]<=pcx_packet_d[110] && pcx_packet_d[122:118]==5'b00000; // Prefetch
	716	cpx_packet_2[144]<=1; // Valid
	717	cpx_packet_2[139]<=0; // L2 miss
	718	cpx_packet_2[138:137]<=0; // Error
	719	cpx_packet_2[136]<=pcx_packet_d[117] \|\| (pcx_packet_d[122:118]==5'b00001) ? 1:0; // Non-cacheble is set on store too
	720	cpx_packet_2[135:134]<=pcx_packet_d[113:112]; // Thread ID
	721	if(pcx_packet_d[122:118]==5'b10000) // IFill
	722	cpx_packet_2[133:131]<={inval_vect1[3],inval_vect1[1:0]};
	723	else
	724	cpx_packet_2[133:131]<=3'b000; // Way valid
	725	cpx_packet_2[130]<=0; // Four byte fill
	726	cpx_packet_2[129]<=pcx_atom_d \|\| (pcx_packet_d[122:118]==5'b00110) \|\| ((pcx_packet_d[122:118]==5'b10000) && !pcx_req_d[4]);
	727	cpx_packet_2[128]<=0; // Prefetch
	728	wb_strobe<=0;
	729	wb_sel<=8'b0;
	730	wb_addr<=64'b0;
	731	wb_data_o<=64'b0;
	732	wb_we<=0;
	733	case(pcx_packet_d[122:118]) // Packet type
	734	5'b00000://Load
	735	begin
	736	cpx_packet_1[143:140]<=4'b0000; // Type
	737	if(!pcx_req_d[4])
	738	begin
	739	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	740	state<=`PCX_REQ_STEP2;
	741	end
	742	else
	743	case(pcx_packet_d[106:104]) //Size
	744	3'b000://Byte
	745	begin
	746	case(pcx_packet_d[64+2:64])
	747	3'b000:cpx_packet_1[127:0]<={wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56],wb_data_i[63:56]};
	748	3'b001:cpx_packet_1[127:0]<={wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48],wb_data_i[55:48]};
	749	3'b010:cpx_packet_1[127:0]<={wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40],wb_data_i[47:40]};
	750	3'b011:cpx_packet_1[127:0]<={wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32],wb_data_i[39:32]};
	751	3'b100:cpx_packet_1[127:0]<={wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24],wb_data_i[31:24]};
	752	3'b101:cpx_packet_1[127:0]<={wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16],wb_data_i[23:16]};
	753	3'b110:cpx_packet_1[127:0]<={wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8],wb_data_i[15: 8]};
	754	3'b111:cpx_packet_1[127:0]<={wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0],wb_data_i[ 7: 0]};
	755	endcase
	756	wb_cycle<=0;
	757	state<=`CPX_READY_1;
	758	end
	759	3'b001://Halfword
	760	begin
	761	case(pcx_packet_d[64+2:64+1])
	762	2'b00:cpx_packet_1[127:0]<={wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48],wb_data_i[63:48]};
	763	2'b01:cpx_packet_1[127:0]<={wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32],wb_data_i[47:32]};
	764	2'b10:cpx_packet_1[127:0]<={wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16],wb_data_i[31:16]};
	765	2'b11:cpx_packet_1[127:0]<={wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0],wb_data_i[15: 0]};
	766	endcase
	767	wb_cycle<=0;
	768	state<=`CPX_READY_1;
	769	end
	770	3'b010://Word
	771	begin
	772	if(pcx_packet_d[64+2]==0)
	773	cpx_packet_1[127:0]<={wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32]};
	774	else
	775	cpx_packet_1[127:0]<={wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0]};
	776	wb_cycle<=0;
	777	state<=`CPX_READY_1;
	778	end
	779	3'b011://Doubleword
	780	begin
	781	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	782	wb_cycle<=0;
	783	state<=`CPX_READY_1;
	784	end
	785	3'b100://Quadword
	786	begin
	787	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	788	wb_cycle<=0;
	789	state<=`CPX_READY_1; // 16 byte access to PROM should just duplicate the data
	790	end
	791	3'b111://Cacheline
	792	begin
	793	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	794	wb_cycle<=0;
	795	state<=`CPX_READY_1; // 16 byte access to PROM should just duplicate the data
	796	end
	797	default:
	798	begin
	799	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	800	wb_cycle<=0;
	801	state<=`PCX_UNKNOWN;
	802	end
	803	endcase
	804	end
	805	5'b00001://Store
	806	begin
	807	cpx_packet_1[143:140]<=4'b0100; // Type
	808	cpx_packet_1[127:0]<={2'b0,pcx_packet_d[109]/BIS/,2'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	809	// if((pcx_packet_d[110:109]==2'b01) && (pcx_packet_d[64+5:64]==0) && !inval_vect0[3] && !inval_vect1[3]) // Block init store
	810	// state<=`PCX_BIS;
	811	// else
	812	// begin
	813	wb_cycle<=0;
	814	state<=`CPX_READY_1;
	815	// end
	816	end
	817	5'b00010://CAS
	818	begin
	819	cpx_packet_1[143:140]<=4'b0000; // Load return for first packet
	820	cpx_packet_2[143:140]<=4'b0100; // Store ACK for second packet
	821	cpx_packet_2[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	822	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	823	state<=`PCX_REQ_STEP2;
	824	end
	825	5'b00100://STRLOAD
	826	begin
	827	cpx_packet_1[143:140]<=4'b0010; // Type
	828	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	829	state<=`PCX_REQ_STEP2;
	830	end
	831	5'b00101://STRSTORE
	832	begin
	833	cpx_packet_1[143:140]<=4'b0110; // Type
	834	cpx_packet_1[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	835	wb_cycle<=0;
	836	state<=`CPX_READY_1;
	837	end
	838	5'b00110://SWAP/LDSTUB
	839	begin
	840	cpx_packet_1[143:140]<=4'b0000; // Load return for first packet
	841	cpx_packet_2[143:140]<=4'b0100; // Store ACK for second packet
	842	cpx_packet_2[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],store_inv_vec};
	843	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	844	state<=`PCX_REQ_STEP2;
	845	end
	846	5'b10000://IFILL
	847	begin
[26]	848	$display("INFO: OS2WB: PCX_REQ_STEP1_1, IFILL, wb_addr = %x wb_data_i= %x",wb_addr, wb_data_i);
	849	$display("INFO: OS2WB: PCX_REQ_STEP1_1, IFILL, cpx_packet_1 = %x %x",wb_data_i,wb_data_i);
[10]	850	cpx_packet_1[143:140]<=4'b0001; // Type
	851	cpx_packet_2[143:140]<=4'b0001; // Type
	852	if(pcx_req_d[4]) // I/O access
	853	begin
	854	if(pcx_packet_d[64+2]==0)
	855	cpx_packet_1[127:0]<={wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32],wb_data_i[63:32]};
	856	else
	857	cpx_packet_1[127:0]<={wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0],wb_data_i[31:0]};
	858	state<=`CPX_READY_1;
	859	wb_cycle<=0;
	860	end
	861	else
	862	begin
	863	cpx_packet_1[127:0]<={wb_data_i,wb_data_i};
	864	state<=`PCX_REQ_STEP2;
	865	end
	866	end
	867	default:
	868	begin
	869	wb_cycle<=0;
	870	state<=`PCX_UNKNOWN;
	871	end
	872	endcase
	873	end
	874	end
	875	`PCX_REQ_STEP2: // IFill, Load/strload, CAS, SWAP, LDSTUB - alwas load
	876	begin
	877	wb_strobe<=1'b1;
	878	if(pcx_packet_d[122:118]==5'b10000)
	879	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b01000};
	880	else
	881	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+4],4'b1000};
	882	wb_sel<=8'b11111111; // It is always full width for subsequent IFill and load accesses
	883	state<=`PCX_REQ_STEP2_1;
	884	end
	885	`PCX_REQ_STEP2_1:
	886	if(wb_ack==1)
	887	begin
	888	wb_strobe<=0;
	889	wb_sel<=8'b0;
	890	wb_addr<=64'b0;
	891	wb_data_o<=64'b0;
	892	wb_we<=0;
	893	cpx_packet_1[63:0]<=wb_data_i;
	894	if((pcx_packet_d[122:118]!=5'b00000) && (pcx_packet_d[122:118]!=5'b00100))
	895	if(pcx_packet_d[122:118]!=5'b00010) // IFill, SWAP
	896	state<=`PCX_REQ_STEP3;
	897	else
	898	state<=`PCX_REQ_CAS_COMPARE; // CAS
	899	else
	900	begin
	901	wb_cycle<=0;
	902	state<=`CPX_READY_1;
	903	end
	904	end
	905	`PCX_REQ_CAS_COMPARE:
	906	begin
	907	cpx_two_packet<=1;
	908	if(pcx_packet_d[106:104]==3'b010) // 32-bit
	909	case(pcx_packet_d[64+3:64+2])
	910	2'b00:state<=cpx_packet_1[127:96]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	911	2'b01:state<=cpx_packet_1[95:64]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	912	2'b10:state<=cpx_packet_1[63:32]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	913	2'b11:state<=cpx_packet_1[31:0]==pcx_packet_d[63:32] ? `PCX_REQ_STEP3:`CPX_READY_1;
	914	endcase
	915	else
	916	if(pcx_packet_d[64+3]==0)
	917	state<=cpx_packet_1[127:64]==pcx_packet_d[63:0] ? `PCX_REQ_STEP3:`CPX_READY_1;
	918	else
	919	state<=cpx_packet_1[63:0]==pcx_packet_d[63:0] ? `PCX_REQ_STEP3:`CPX_READY_1;
	920	end
	921	`PCX_REQ_STEP3: // 256-bit IFILL; CAS, SWAP and LDSTUB store
	922	begin
	923	if(pcx_packet_d[122:118]==5'b10000)
	924	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b10000};
	925	else
	926	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+3],3'b000}; // CAS or SWAP save
	927	cpx_two_packet<=1;
	928	if(pcx_packet_d[122:118]==5'b10000)
	929	wb_we<=0;
	930	else
	931	wb_we<=1;
	932	wb_strobe<=1'b1;
	933	if(pcx_packet_d[122:118]==5'b00010) // CAS
	934	if(pcx_packet_d[106:104]==3'b010)
	935	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111;
	936	else
	937	wb_sel<=8'b11111111; //CASX
	938	else
	939	if(pcx_packet_d[122:118]==5'b00110) //SWAP or LDSTUB
	940	if(pcx_packet_d[106:104]==3'b000) //LDSTUB
	941	case(pcx_packet_d[64+2:64])
	942	3'b000:wb_sel<=8'b10000000;
	943	3'b001:wb_sel<=8'b01000000;
	944	3'b010:wb_sel<=8'b00100000;
	945	3'b011:wb_sel<=8'b00010000;
	946	3'b100:wb_sel<=8'b00001000;
	947	3'b101:wb_sel<=8'b00000100;
	948	3'b110:wb_sel<=8'b00000010;
	949	3'b111:wb_sel<=8'b00000001;
	950	endcase
	951	else
	952	wb_sel<=(pcx_packet_d[64+2]==0) ? 8'b11110000:8'b00001111; ///SWAP is always 32-bit
	953	else
	954	wb_sel<=8'b11111111; // It is always full width for subsequent IFill accesses
	955	if(pcx_packet_d[122:118]==5'b00110) //SWAP or LDSTUB
	956	wb_data_o<={pcx_packet_d[63:32],pcx_packet_d[63:32]};
	957	// wb_data_o<=pcx_packet_d[63:0];
	958	else
	959	wb_data_o<=pcx_packet_2nd[63:0]; // CAS store second packet data
	960	// if(pcx_packet_d[106:104]==3'b010)
	961	// wb_data_o<={pcx_packet_2nd[63:32],pcx_packet_2nd[63:32]}; // CAS store second packet data
	962	// else
	963	// wb_data_o<=pcx_packet_2nd[63:0];
	964	state<=`PCX_REQ_STEP3_1;
	965	end
	966	`PCX_REQ_STEP3_1:
	967	if(wb_ack==1)
	968	begin
	969	wb_strobe<=0;
	970	wb_sel<=8'b0;
	971	wb_addr<=64'b0;
	972	wb_we<=0;
	973	wb_data_o<=64'b0;
	974	if(pcx_packet_d[122:118]==5'b10000) // IFill
	975	begin
	976	cpx_packet_2[127:64]<=wb_data_i;
	977	state<=`PCX_REQ_STEP4;
	978	end
	979	else
	980	begin
	981	wb_cycle<=0;
	982	state<=`CPX_READY_1;
	983	end
	984	end
	985	`PCX_REQ_STEP4: // 256-bit IFILL only
	986	begin
	987	wb_strobe<=1'b1;
	988	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+5],5'b11000};
	989	wb_sel<=8'b11111111; // It is always full width for subsequent accesses
	990	state<=`PCX_REQ_STEP4_1;
	991	end
	992	`PCX_REQ_STEP4_1:
	993	if(wb_ack==1)
	994	begin
	995	wb_cycle<=0;
	996	wb_strobe<=0;
	997	wb_sel<=8'b0;
	998	wb_addr<=64'b0;
	999	wb_we<=0;
	1000	cpx_packet_2[63:0]<=wb_data_i;
	1001	state<=`CPX_READY_1;
	1002	end
	1003	`PCX_BIS: // Block init store
	1004	begin
	1005	wb_strobe<=1'b1;
	1006	wb_we<=1;
	1007	wb_addr<={pcx_req_d,19'b0,pcx_packet_d[103:64+6],6'b001000};
	1008	wb_sel<=8'b11111111;
	1009	wb_data_o<=64'b0;
	1010	state<=`PCX_BIS_1;
	1011	end
	1012	`PCX_BIS_1:
	1013	if(wb_ack)
	1014	begin
	1015	wb_strobe<=0;
	1016	if(wb_addr[39:0]<(pcx_packet_d[64+39:64]+8*7))
	1017	state<=`PCX_BIS_2;
	1018	else
	1019	begin
	1020	wb_cycle<=0;
	1021	wb_sel<=0;
	1022	wb_we<=0;
	1023	wb_addr<=64'b0;
	1024	state<=`CPX_READY_1;
	1025	end
	1026	end
	1027	`PCX_BIS_2:
	1028	begin
	1029	wb_strobe<=1'b1;
	1030	wb_addr[5:0]<=wb_addr[5:0]+8;
	1031	state<=`PCX_BIS_1;
	1032	end
	1033	`PCX_FP_1:
	1034	begin
	1035	fp_pcx<=pcx_packet_d;
	1036	fp_req<=1;
	1037	state<=`PCX_FP_2;
	1038	if(`DEBUGGING)
	1039	begin
	1040	wb_addr<=pcx_packet_d[103:64];
	1041	wb_data_o<=pcx_packet_d[63:0];
	1042	wb_sel<=8'h22;
	1043	end
	1044	end
	1045	`PCX_FP_2:
	1046	begin
	1047	fp_pcx<=pcx_packet_2nd;
	1048	state<=`FP_WAIT;
	1049	if(`DEBUGGING)
	1050	begin
	1051	wb_addr<=pcx_packet_2nd[103:64];
	1052	wb_data_o<=pcx_packet_d[63:0];
	1053	wb_sel<=8'h23;
	1054	end
	1055	end
	1056	`FP_WAIT:
	1057	begin
	1058	fp_pcx<=124'b0;
	1059	fp_req<=0;
	1060	if(fp_rdy)
	1061	state<=`CPX_FP;
	1062	if(`DEBUGGING)
	1063	wb_sel<=8'h24;
	1064	end
	1065	`CPX_FP:
	1066	if(fp_cpx[144]) // Packet valid
	1067	begin
	1068	cpx_packet_1<=fp_cpx;
	1069	state<=`CPX_READY_1;
	1070	if(`DEBUGGING)
	1071	begin
	1072	wb_addr<=fp_cpx[63:0];
	1073	wb_data_o<=fp_cpx[127:64];
	1074	end
	1075	end
	1076	else
	1077	if(!fp_rdy)
	1078	state<=`FP_WAIT; // Else wait for another one if it is not here still
	1079	`CPX_SEND_ETH_IRQ:
	1080	begin
	1081	cpx_packet_1<=145'h1_7_000_000000000000001D_000000000000_001D;
	1082	eth_int_sent<=0;
	1083	state<=`CPX_READY_1;
	1084	end
	1085	`CPX_INT_VEC_DIS:
	1086	begin
	1087	if(pcx_packet_d[12:10]==3'b000)
	1088	cpx_two_packet<=1; // Send interrupt only if it is for this core
	1089	cpx_packet_1[144:140]<=5'b10100;
	1090	cpx_packet_1[139:137]<=0;
	1091	cpx_packet_1[136]<=1;
	1092	cpx_packet_1[135:134]<=pcx_packet_d[113:112]; // Thread ID
	1093	cpx_packet_1[133:130]<=0;
	1094	cpx_packet_1[129]<=pcx_atom_d;
	1095	cpx_packet_1[128]<=0;
	1096	cpx_packet_1[127:0]<={5'b0,pcx_packet_d[64+5:64+4],3'b0,pcx_packet_d[64+11:64+6],112'b0};
	1097	cpx_packet_2<={9'h170,54'h0,pcx_packet_d[17:0],46'h0,pcx_packet_d[17:0]};
	1098	state<=`CPX_READY_1;
	1099	end
	1100	`CPX_READY_1:
	1101	begin
[26]	1102	$display("INFO: OS2WB: CPX_READY_1");
[10]	1103	cpx_ready<=1;
	1104	cpx_packet<=cpx_packet_1;
	1105	cnt<=cnt+1;
	1106	if(`DEBUGGING)
	1107	if(multi_hit \|\| multi_hit1)
	1108	wb_sel<=8'h11;
	1109	if(!cpx_two_packet)
	1110	state<=`PCX_IDLE;
	1111	else
	1112	//if(cnt==4'b1111 \|\| pcx_packet_d[103:64]!=40'h9800000800)
	1113	state<=`CPX_READY_2;
	1114	end
	1115	`CPX_READY_2:
	1116	begin
[26]	1117	$display("INFO: OS2WB: CPX_READY_2");
[10]	1118	cpx_ready<=1;
	1119	cpx_packet<=cpx_packet_2;
	1120	state<=`PCX_IDLE;
	1121	end
	1122	`PCX_UNKNOWN:
	1123	begin
[26]	1124	$display("INFO: OS2WB: PCX_UNKNOWN");
[10]	1125	wb_sel<=8'b10100101; // Illegal eye-catching value for debugging
	1126	state<=`PCX_IDLE;
	1127	end
	1128	endcase
	1129
	1130	/* Cache directory checking:
	1131	Load: allocate D if cacheable, check I, invalidate&deallocate if found
	1132	Store: check I, invalidate&deallocate if found; check D, invalidate if found
	1133	IFill: allocate I if cacheable, check D, invalidate&deallocate if found
	1134	SWAP/LDSTUB: check I, invalidate&deallocate if found; check D, invalidate&deallocate if found
	1135	CAS: Like SWAP
	1136
	1137	Allocation and querying is made simultaneously at GOT_PCX_REQ
	1138	(memory read mode does not matter as long as allocation and invalidation
	1139	are never made to the same directory, so if memory is written its output will not be checked)
	1140	Invalidation vectors are built during PCX_REQ_STEP1, or Invalidate all ways issued
	1141	During PCX_REQ_STEP1_1 directory is deallocated if needed
	1142
	1143	*/
	1144
	1145	// Directory enable
	1146	assign dir_en=((state==`GOT_PCX_REQ) \|\| (state==`PCX_REQ_STEP1) \|\| cache_init \|\|
	1147	((state==`PCX_REQ_STEP1_1) && wb_ack)) ? 1:0;
	1148
	1149	// ICache deallocation flag
	1150	assign loadstore=((pcx_packet_d[122:118]==5'b00000) && !pcx_packet_d[117] && !pcx_packet_d[110]) \|\| // cacheable load, not prefetch
	1151	(pcx_packet_d[122:118]==5'b00001) \|\| (pcx_packet_d[122:118]==5'b00010) \|\| // Store, CAS
	1152	(pcx_packet_d[122:118]==5'b00110) \|\| (pcx_packet_d[122:118]==5'b00101); // SWAP/LDSTUB, StrStore
	1153
	1154	// DCache deallocation flag
	1155	assign ifillcas=(pcx_packet_d[122:118]==5'b00110) \|\| (pcx_packet_d[122:118]==5'b00010) \|\| //SWAP, CAS
	1156	(pcx_packet_d[122:118]==5'b10000) \|\| (pcx_packet_d[122:118]==5'b00101) \|\| // IFill, StrStore
	1157	((pcx_packet_d[122:118]==5'b00001) && pcx_packet_d[110:109]!=2'b00); // Block (or init) store
	1158
	1159	// DCache allocation flag
	1160	assign cacheload=(pcx_packet[122:118]==5'b00000) && !pcx_packet[110] && !pcx_packet[117] && !pcx_packet[111];
	1161
	1162	// ICache allocation flag
	1163	assign cacheifill=(pcx_packet[122:118]==5'b10000) && !pcx_packet[117] && !pcx_packet[111];
	1164
	1165	assign dcache0_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b00) && cacheload;
	1166	assign dcache0_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_00) && ifillcas;
	1167	assign dcache0_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_00) && ifillcas;
	1168
	1169	assign dcache1_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b01) && cacheload;
	1170	assign dcache1_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_01) && ifillcas;
	1171	assign dcache1_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_01) && ifillcas;
	1172
	1173	assign dcache2_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b10) && cacheload;
	1174	assign dcache2_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_10) && ifillcas;
	1175	assign dcache2_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_10) && ifillcas;
	1176
	1177	assign dcache3_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b11) && cacheload;
	1178	assign dcache3_dealloc0=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_1_11) && ifillcas;
	1179	assign dcache3_dealloc1=(state==`PCX_REQ_STEP1_1) && (inval_vect1==4'b1_1_11) && ifillcas;
	1180
	1181	assign icache0_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b00) && cacheifill;
	1182	assign icache0_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_00) && loadstore;
	1183
	1184	assign icache1_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b01) && cacheifill;
	1185	assign icache1_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_01) && loadstore;
	1186
	1187	assign icache2_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b10) && cacheifill;
	1188	assign icache2_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_10) && loadstore;
	1189
	1190	assign icache3_alloc=(state==`GOT_PCX_REQ) && (pcx_packet[108:107]==2'b11) && cacheifill;
	1191	assign icache3_dealloc=(state==`PCX_REQ_STEP1_1) && (inval_vect0==4'b1_0_11) && loadstore;
	1192
	1193	assign dcache_inval_all=(state==`PCX_REQ_STEP1) && pcx_packet_d[111] && pcx_packet_d[122:118]==5'b00000;
	1194	assign icache_inval_all=(state==`PCX_REQ_STEP1) && pcx_packet_d[111] && pcx_packet_d[122:118]==5'b10000;
	1195
	1196	`define INVAL_TAG 29'h10000000
	1197
	1198	// DCache least address bit for first bank
	1199	// it should be 0 for IFill (1 is hardcoded for second bank)
	1200	assign dcache_la=(state==`GOT_PCX_REQ) ? (pcx_packet[122:118]==5'b10000 ? 1'b0:pcx_packet[64+4]):
	1201	(pcx_packet_d[122:118]==5'b10000 ? 1'b0:pcx_packet_d[64+4]);
	1202
	1203	wire [ 6:0] dcache_index;
	1204	wire [28:0] dcache_data;
	1205	assign dcache_index=(state==`GOT_PCX_REQ) ? pcx_packet[64+10:64+5]:pcx_packet_d[64+10:64+5];
	1206	assign dcache_data=(state==`GOT_PCX_REQ) ? pcx_packet[64+39:64+11]:`INVAL_TAG;
	1207
	1208	cachedir dcache0 (
	1209	.clock(clk),
	1210	.enable(dir_en),
	1211	.wren_a(dcache0_alloc \|\| dcache0_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1212	.address_a({1'b0,dcache_index,dcache_la}),
	1213	.data_a(dcache_data),
	1214	.q_a(dcache0_do0),
	1215
	1216	.wren_b(dcache0_dealloc1),
	1217	.address_b({1'b0,dcache_index,1'b1}),
	1218	.data_b(`INVAL_TAG),
	1219	.q_b(dcache0_do1)
	1220	);
	1221
	1222	cachedir dcache1 (
	1223	.clock(clk),
	1224	.enable(dir_en),
	1225	.wren_a(dcache1_alloc \|\| dcache1_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1226	.address_a({1'b0,dcache_index,dcache_la}),
	1227	.data_a(dcache_data),
	1228	.q_a(dcache1_do0),
	1229
	1230	.wren_b(dcache1_dealloc1),
	1231	.address_b({1'b0,dcache_index,1'b1}),
	1232	.data_b(`INVAL_TAG),
	1233	.q_b(dcache1_do1)
	1234	);
	1235
	1236	cachedir dcache2 (
	1237	.clock(clk),
	1238	.enable(dir_en),
	1239	.wren_a(dcache2_alloc \|\| dcache2_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1240	.address_a({1'b0,dcache_index,dcache_la}),
	1241	.data_a(dcache_data),
	1242	.q_a(dcache2_do0),
	1243
	1244	.wren_b(dcache2_dealloc1),
	1245	.address_b({1'b0,dcache_index,1'b1}),
	1246	.data_b(`INVAL_TAG),
	1247	.q_b(dcache2_do1)
	1248	);
	1249
	1250	cachedir dcache3 (
	1251	.clock(clk),
	1252	.enable(dir_en),
	1253	.wren_a(dcache3_alloc \|\| dcache3_dealloc0 \|\| dcache_inval_all \|\| cache_init),
	1254	.address_a({1'b0,dcache_index,dcache_la}),
	1255	.data_a(dcache_data),
	1256	.q_a(dcache3_do0),
	1257
	1258	.wren_b(dcache3_dealloc1),
	1259	.address_b({1'b0,dcache_index,1'b1}),
	1260	.data_b(`INVAL_TAG),
	1261	.q_b(dcache3_do1)
	1262	);
	1263
	1264	assign dcache0_hit={dcache3_do0==pcx_packet_d[64+39:64+11],
	1265	dcache2_do0==pcx_packet_d[64+39:64+11],
	1266	dcache1_do0==pcx_packet_d[64+39:64+11],
	1267	dcache0_do0==pcx_packet_d[64+39:64+11]};
	1268	assign dcache1_hit={dcache3_do1==pcx_packet_d[64+39:64+11],
	1269	dcache2_do1==pcx_packet_d[64+39:64+11],
	1270	dcache1_do1==pcx_packet_d[64+39:64+11],
	1271	dcache0_do1==pcx_packet_d[64+39:64+11]};
	1272
	1273	wire [ 6:0] icache_index;
	1274	wire [28:0] icache_data;
	1275	assign icache_index=(state==`GOT_PCX_REQ) ? pcx_packet[64+11:64+5]:pcx_packet_d[64+11:64+5];
	1276	assign icache_data=(state==`GOT_PCX_REQ) ? {pcx_packet[64+39:64+12],1'b0}:`INVAL_TAG;
	1277
	1278	cachedir icache01 (
	1279	.clock(clk),
	1280	.enable(dir_en),
	1281	.wren_a(icache0_alloc \|\| icache0_dealloc \|\| icache_inval_all \|\| cache_init),
[26]	1282	.address_a({2'b00,icache_index}),
[10]	1283	.data_a(icache_data),
	1284	.q_a(icache0_do),
	1285
	1286	.wren_b(icache1_alloc \|\| icache1_dealloc \|\| icache_inval_all \|\| cache_init),
[26]	1287	.address_b({2'b01,icache_index}),
[10]	1288	.data_b(icache_data),
	1289	.q_b(icache1_do)
	1290	);
	1291
	1292	cachedir icache23 (
	1293	.clock(clk),
	1294	.enable(dir_en),
	1295	.wren_a(icache2_alloc \|\| icache2_dealloc \|\| icache_inval_all \|\| cache_init),
[26]	1296	.address_a({2'b00,icache_index}),
[10]	1297	.data_a(icache_data),
	1298	.q_a(icache2_do),
	1299
	1300	.wren_b(icache3_alloc \|\| icache3_dealloc \|\| icache_inval_all \|\| cache_init),
[26]	1301	.address_b({2'b01,icache_index}),
[10]	1302	.data_b(icache_data),
	1303	.q_b(icache3_do)
	1304	);
	1305
	1306	assign icache_hit={icache3_do[28:1]==pcx_packet_d[64+39:64+12],
	1307	icache2_do[28:1]==pcx_packet_d[64+39:64+12],
	1308	icache1_do[28:1]==pcx_packet_d[64+39:64+12],
	1309	icache0_do[28:1]==pcx_packet_d[64+39:64+12]};
	1310
	1311	/*
	1312	case(pcx_packet_d[122:118]) // Packet type
	1313	5'b00000://Load
	1314	5'b00001://Store
	1315	5'b00010://CAS
	1316	5'b00100://STRLOAD
	1317	5'b00101://STRSTORE
	1318	5'b00110://SWAP
	1319	5'b01001://INT
	1320	//5'b01010://FP1
	1321	//5'b01011://FP2
	1322	//5'b01101://FWDREQ
	1323	//5'b01110://FWDREPL
	1324	5'b10000://IFILL
	1325	endcase
	1326	*/
[14]	1327	endmodule

Note: See TracBrowser for help on using the repository browser.

Context Navigation

source: XOpenSparcT1/trunk/os2wb/os2wb.v @ 36

Download in other formats: