| 1 |
3 |
dgisselq |
////////////////////////////////////////////////////////////////////////////////
|
| 2 |
|
|
//
|
| 3 |
|
|
// Filename: pipefetch.v
|
| 4 |
|
|
//
|
| 5 |
|
|
// Project: Zip CPU -- a small, lightweight, RISC CPU soft core
|
| 6 |
|
|
//
|
| 7 |
|
|
// Purpose: Keeping our CPU fed with instructions, at one per clock and
|
| 8 |
|
|
// with no stalls, can be quite a chore. Worse, the Wishbone
|
| 9 |
|
|
// takes a couple of cycles just to read one instruction from
|
| 10 |
|
|
// the bus. However, if we use pipeline accesses to the Wishbone
|
| 11 |
|
|
// bus, then we can read more and faster. Further, if we cache
|
| 12 |
|
|
// these results so that we have them before we need them, then
|
| 13 |
|
|
// we have a chance of keeping our CPU from stalling. Those are
|
| 14 |
|
|
// the purposes of this instruction fetch module: 1) Pipeline
|
| 15 |
|
|
// wishbone accesses, and 2) an instruction cache.
|
| 16 |
|
|
//
|
| 17 |
|
|
// 20150919 -- Fixed a nasty race condition whereby the pipefetch routine
|
| 18 |
|
|
// would produce either the same instruction twice, or skip
|
| 19 |
|
|
// an instruction. This condition was dependent on the CPU stall
|
| 20 |
|
|
// condition, and would only take place if the pipeline wasn't
|
| 21 |
|
|
// completely full throughout the stall.
|
| 22 |
|
|
//
|
| 23 |
|
|
// Interface support was also added for trapping on illegal
|
| 24 |
|
|
// instructions (i.e., instruction fetches that cause bus errors),
|
| 25 |
|
|
// however the internal interface has not caught up to supporting
|
| 26 |
|
|
// these exceptions yet.
|
| 27 |
|
|
//
|
| 28 |
|
|
// Creator: Dan Gisselquist, Ph.D.
|
| 29 |
|
|
// Gisselquist Technology, LLC
|
| 30 |
|
|
//
|
| 31 |
|
|
////////////////////////////////////////////////////////////////////////////////
|
| 32 |
|
|
//
|
| 33 |
|
|
// Copyright (C) 2015, Gisselquist Technology, LLC
|
| 34 |
|
|
//
|
| 35 |
|
|
// This program is free software (firmware): you can redistribute it and/or
|
| 36 |
|
|
// modify it under the terms of the GNU General Public License as published
|
| 37 |
|
|
// by the Free Software Foundation, either version 3 of the License, or (at
|
| 38 |
|
|
// your option) any later version.
|
| 39 |
|
|
//
|
| 40 |
|
|
// This program is distributed in the hope that it will be useful, but WITHOUT
|
| 41 |
|
|
// ANY WARRANTY; without even the implied warranty of MERCHANTIBILITY or
|
| 42 |
|
|
// FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
| 43 |
|
|
// for more details.
|
| 44 |
|
|
//
|
| 45 |
|
|
// License: GPL, v3, as defined and found on www.gnu.org,
|
| 46 |
|
|
// http://www.gnu.org/licenses/gpl.html
|
| 47 |
|
|
//
|
| 48 |
|
|
//
|
| 49 |
|
|
////////////////////////////////////////////////////////////////////////////////
|
| 50 |
|
|
//
|
| 51 |
|
|
module pipefetch(i_clk, i_rst, i_new_pc, i_clear_cache, i_stall_n, i_pc,
|
| 52 |
|
|
o_i, o_pc, o_v,
|
| 53 |
|
|
o_wb_cyc, o_wb_stb, o_wb_we, o_wb_addr, o_wb_data,
|
| 54 |
|
|
i_wb_ack, i_wb_stall, i_wb_err, i_wb_data, i_wb_request,
|
| 55 |
|
|
o_illegal);
|
| 56 |
|
|
parameter RESET_ADDRESS=32'h0010_0000,
|
| 57 |
|
|
LGCACHELEN = 6, ADDRESS_WIDTH=24,
|
| 58 |
|
|
CACHELEN=(1<<LGCACHELEN), BUSW=32, AW=ADDRESS_WIDTH;
|
| 59 |
|
|
input i_clk, i_rst, i_new_pc,
|
| 60 |
|
|
i_clear_cache, i_stall_n;
|
| 61 |
|
|
input [(AW-1):0] i_pc;
|
| 62 |
|
|
output reg [(BUSW-1):0] o_i;
|
| 63 |
|
|
output reg [(AW-1):0] o_pc;
|
| 64 |
|
|
output wire o_v;
|
| 65 |
|
|
//
|
| 66 |
|
|
output reg o_wb_cyc, o_wb_stb;
|
| 67 |
|
|
output wire o_wb_we;
|
| 68 |
|
|
output reg [(AW-1):0] o_wb_addr;
|
| 69 |
|
|
output wire [(BUSW-1):0] o_wb_data;
|
| 70 |
|
|
//
|
| 71 |
|
|
input i_wb_ack, i_wb_stall, i_wb_err;
|
| 72 |
|
|
input [(BUSW-1):0] i_wb_data;
|
| 73 |
|
|
//
|
| 74 |
|
|
// Is the (data) memory unit also requesting access to the bus?
|
| 75 |
|
|
input i_wb_request;
|
| 76 |
|
|
output wire o_illegal;
|
| 77 |
|
|
|
| 78 |
|
|
// Fixed bus outputs: we read from the bus only, never write.
|
| 79 |
|
|
// Thus the output data is ... irrelevant and don't care. We set it
|
| 80 |
|
|
// to zero just to set it to something.
|
| 81 |
|
|
assign o_wb_we = 1'b0;
|
| 82 |
|
|
assign o_wb_data = 0;
|
| 83 |
|
|
|
| 84 |
|
|
reg [(AW-1):0] r_cache_base;
|
| 85 |
|
|
reg [(LGCACHELEN):0] r_nvalid, r_acks_waiting;
|
| 86 |
|
|
reg [(BUSW-1):0] cache[0:(CACHELEN-1)];
|
| 87 |
|
|
|
| 88 |
|
|
wire [(LGCACHELEN-1):0] w_cache_offset;
|
| 89 |
|
|
reg [1:0] r_cache_offset;
|
| 90 |
|
|
|
| 91 |
|
|
reg r_addr_set;
|
| 92 |
|
|
reg [(AW-1):0] r_addr;
|
| 93 |
|
|
|
| 94 |
|
|
wire [(AW-1):0] bus_nvalid;
|
| 95 |
|
|
assign bus_nvalid = { {(AW-LGCACHELEN-1){1'b0}}, r_nvalid };
|
| 96 |
|
|
|
| 97 |
|
|
// What are some of the conditions for which we need to restart the
|
| 98 |
|
|
// cache?
|
| 99 |
|
|
wire w_pc_out_of_bounds;
|
| 100 |
|
|
assign w_pc_out_of_bounds = ((i_new_pc)&&((r_nvalid == 0)
|
| 101 |
|
|
||(i_pc < r_cache_base)
|
| 102 |
|
|
||(i_pc >= r_cache_base + CACHELEN)
|
| 103 |
|
|
||(i_pc >= r_cache_base + bus_nvalid+5)));
|
| 104 |
|
|
wire w_ran_off_end_of_cache;
|
| 105 |
|
|
assign w_ran_off_end_of_cache =((r_addr_set)&&((r_addr < r_cache_base)
|
| 106 |
|
|
||(r_addr >= r_cache_base + CACHELEN)
|
| 107 |
|
|
||(r_addr >= r_cache_base + bus_nvalid+5)));
|
| 108 |
|
|
wire w_running_out_of_cache;
|
| 109 |
|
|
assign w_running_out_of_cache = (r_addr_set)
|
| 110 |
|
|
&&(r_addr >= r_cache_base +
|
| 111 |
|
|
// {{(AW-LGCACHELEN-1),{1'b0}},2'b11,
|
| 112 |
|
|
// {(LGCACHELEN-1){1'b0}}})
|
| 113 |
|
|
// (1<<(LGCACHELEN-2)) + (1<<(LGCACHELEN-1)))
|
| 114 |
|
|
+(3<<(LGCACHELEN-2)))
|
| 115 |
|
|
&&(|r_nvalid[(LGCACHELEN):(LGCACHELEN-1)]);
|
| 116 |
|
|
|
| 117 |
|
|
initial r_cache_base = RESET_ADDRESS;
|
| 118 |
|
|
always @(posedge i_clk)
|
| 119 |
|
|
begin
|
| 120 |
|
|
if ((i_rst)||(i_clear_cache)||((o_wb_cyc)&&(i_wb_err)))
|
| 121 |
|
|
begin
|
| 122 |
|
|
o_wb_cyc <= 1'b0;
|
| 123 |
|
|
o_wb_stb <= 1'b0;
|
| 124 |
|
|
// r_cache_base <= RESET_ADDRESS;
|
| 125 |
|
|
// end else if ((~o_wb_cyc)&&(i_new_pc)&&(r_nvalid != 0)
|
| 126 |
|
|
// &&(i_pc >= r_cache_base)
|
| 127 |
|
|
// &&(i_pc < r_cache_base + bus_nvalid))
|
| 128 |
|
|
// begin
|
| 129 |
|
|
// The new instruction is in our cache, do nothing
|
| 130 |
|
|
// with the bus here.
|
| 131 |
|
|
end else if ((o_wb_cyc)&&(w_pc_out_of_bounds))
|
| 132 |
|
|
begin
|
| 133 |
|
|
// We need to abandon our bus action to start over in
|
| 134 |
|
|
// a new region, setting up a new cache. This may
|
| 135 |
|
|
// happen mid cycle while waiting for a result. By
|
| 136 |
|
|
// dropping o_wb_cyc, we state that we are no longer
|
| 137 |
|
|
// interested in that result--whatever it might be.
|
| 138 |
|
|
o_wb_cyc <= 1'b0;
|
| 139 |
|
|
o_wb_stb <= 1'b0;
|
| 140 |
|
|
end else if ((~o_wb_cyc)&&(~r_nvalid[LGCACHELEN])&&(~i_wb_request)&&(r_addr_set))
|
| 141 |
|
|
begin
|
| 142 |
|
|
// Restart a bus cycle that was interrupted when the
|
| 143 |
|
|
// data section wanted access to our bus.
|
| 144 |
|
|
o_wb_cyc <= 1'b1;
|
| 145 |
|
|
o_wb_stb <= 1'b1;
|
| 146 |
|
|
// o_wb_addr <= r_cache_base + bus_nvalid;
|
| 147 |
|
|
end else if ((~o_wb_cyc)&&(
|
| 148 |
|
|
(w_pc_out_of_bounds)||(w_ran_off_end_of_cache)))
|
| 149 |
|
|
begin
|
| 150 |
|
|
// Start a bus transaction
|
| 151 |
|
|
o_wb_cyc <= 1'b1;
|
| 152 |
|
|
o_wb_stb <= 1'b1;
|
| 153 |
|
|
// o_wb_addr <= (i_new_pc) ? i_pc : r_addr;
|
| 154 |
|
|
// r_nvalid <= 0;
|
| 155 |
|
|
// r_cache_base <= (i_new_pc) ? i_pc : r_addr;
|
| 156 |
|
|
// w_cache_offset <= 0;
|
| 157 |
|
|
end else if ((~o_wb_cyc)&&(w_running_out_of_cache))
|
| 158 |
|
|
begin
|
| 159 |
|
|
// If we're using the last quarter of the cache, then
|
| 160 |
|
|
// let's start a bus transaction to extend the cache.
|
| 161 |
|
|
o_wb_cyc <= 1'b1;
|
| 162 |
|
|
o_wb_stb <= 1'b1;
|
| 163 |
|
|
// o_wb_addr <= r_cache_base + (1<<(LGCACHELEN));
|
| 164 |
|
|
// r_nvalid <= r_nvalid - (1<<(LGCACHELEN-2));
|
| 165 |
|
|
// r_cache_base <= r_cache_base + (1<<(LGCACHELEN-2));
|
| 166 |
|
|
// w_cache_offset <= w_cache_offset + (1<<(LGCACHELEN-2));
|
| 167 |
|
|
end else if (o_wb_cyc)
|
| 168 |
|
|
begin
|
| 169 |
|
|
// This handles everything ... but the case where
|
| 170 |
|
|
// while reading we need to extend our cache.
|
| 171 |
|
|
if ((o_wb_stb)&&(~i_wb_stall))
|
| 172 |
|
|
begin
|
| 173 |
|
|
// o_wb_addr <= o_wb_addr + 1;
|
| 174 |
|
|
if ((o_wb_addr - r_cache_base >= CACHELEN-1)
|
| 175 |
|
|
||(i_wb_request))
|
| 176 |
|
|
o_wb_stb <= 1'b0;
|
| 177 |
|
|
end
|
| 178 |
|
|
|
| 179 |
|
|
if (i_wb_ack)
|
| 180 |
|
|
begin
|
| 181 |
|
|
// r_nvalid <= r_nvalid + 1;
|
| 182 |
|
|
if ((r_acks_waiting == 1)&&(~o_wb_stb))
|
| 183 |
|
|
o_wb_cyc <= 1'b0;
|
| 184 |
|
|
end else if ((r_acks_waiting == 0)&&(~o_wb_stb))
|
| 185 |
|
|
o_wb_cyc <= 1'b0;
|
| 186 |
|
|
end
|
| 187 |
|
|
end
|
| 188 |
|
|
|
| 189 |
|
|
|
| 190 |
|
|
initial r_nvalid = 0;
|
| 191 |
|
|
always @(posedge i_clk)
|
| 192 |
|
|
if ((i_rst)||(i_clear_cache)) // Required, so we can reload memoy and then reset
|
| 193 |
|
|
r_nvalid <= 0;
|
| 194 |
|
|
else if ((~o_wb_cyc)&&(
|
| 195 |
|
|
(w_pc_out_of_bounds)||(w_ran_off_end_of_cache)))
|
| 196 |
|
|
r_nvalid <= 0;
|
| 197 |
|
|
else if ((~o_wb_cyc)&&(w_running_out_of_cache))
|
| 198 |
|
|
r_nvalid[LGCACHELEN:(LGCACHELEN-2)]
|
| 199 |
|
|
<= r_nvalid[LGCACHELEN:(LGCACHELEN-2)] +3'b111;
|
| 200 |
|
|
// i.e. - (1<<(LGCACHELEN-2));
|
| 201 |
|
|
else if ((o_wb_cyc)&&(i_wb_ack))
|
| 202 |
|
|
r_nvalid <= r_nvalid + {{(LGCACHELEN){1'b0}},1'b1}; // +1;
|
| 203 |
|
|
|
| 204 |
|
|
always @(posedge i_clk)
|
| 205 |
|
|
if (i_clear_cache)
|
| 206 |
|
|
r_cache_base <= i_pc;
|
| 207 |
|
|
else if ((~o_wb_cyc)&&(
|
| 208 |
|
|
(w_pc_out_of_bounds)
|
| 209 |
|
|
||(w_ran_off_end_of_cache)))
|
| 210 |
|
|
r_cache_base <= (i_new_pc) ? i_pc : r_addr;
|
| 211 |
|
|
else if ((~o_wb_cyc)&&(w_running_out_of_cache))
|
| 212 |
|
|
r_cache_base[(AW-1):(LGCACHELEN-2)]
|
| 213 |
|
|
<= r_cache_base[(AW-1):(LGCACHELEN-2)]
|
| 214 |
|
|
+ {{(AW-LGCACHELEN+1){1'b0}},1'b1};
|
| 215 |
|
|
// i.e. + (1<<(LGCACHELEN-2));
|
| 216 |
|
|
|
| 217 |
|
|
always @(posedge i_clk)
|
| 218 |
|
|
if (i_clear_cache)
|
| 219 |
|
|
r_cache_offset <= 0;
|
| 220 |
|
|
else if ((~o_wb_cyc)&&(
|
| 221 |
|
|
(w_pc_out_of_bounds)
|
| 222 |
|
|
||(w_ran_off_end_of_cache)))
|
| 223 |
|
|
r_cache_offset <= 0;
|
| 224 |
|
|
else if ((~o_wb_cyc)&&(w_running_out_of_cache))
|
| 225 |
|
|
r_cache_offset[1:0] <= r_cache_offset[1:0] + 2'b01;
|
| 226 |
|
|
assign w_cache_offset = { r_cache_offset, {(LGCACHELEN-2){1'b0}} };
|
| 227 |
|
|
|
| 228 |
|
|
always @(posedge i_clk)
|
| 229 |
|
|
if (i_clear_cache)
|
| 230 |
|
|
o_wb_addr <= i_pc;
|
| 231 |
|
|
else if ((o_wb_cyc)&&(w_pc_out_of_bounds))
|
| 232 |
|
|
begin
|
| 233 |
|
|
if (i_wb_ack)
|
| 234 |
|
|
o_wb_addr <= r_cache_base + bus_nvalid+1;
|
| 235 |
|
|
else
|
| 236 |
|
|
o_wb_addr <= r_cache_base + bus_nvalid;
|
| 237 |
|
|
end else if ((~o_wb_cyc)&&((w_pc_out_of_bounds)
|
| 238 |
|
|
||(w_ran_off_end_of_cache)))
|
| 239 |
|
|
o_wb_addr <= (i_new_pc) ? i_pc : r_addr;
|
| 240 |
|
|
else if ((o_wb_stb)&&(~i_wb_stall)) // && o_wb_cyc
|
| 241 |
|
|
o_wb_addr <= o_wb_addr + 1;
|
| 242 |
|
|
|
| 243 |
|
|
initial r_acks_waiting = 0;
|
| 244 |
|
|
always @(posedge i_clk)
|
| 245 |
|
|
if (~o_wb_cyc)
|
| 246 |
|
|
r_acks_waiting <= 0;
|
| 247 |
|
|
// o_wb_cyc *must* be true for all following
|
| 248 |
|
|
else if ((o_wb_stb)&&(~i_wb_stall)&&(~i_wb_ack)) //&&(o_wb_cyc)
|
| 249 |
|
|
r_acks_waiting <= r_acks_waiting + {{(LGCACHELEN){1'b0}},1'b1};
|
| 250 |
|
|
else if ((i_wb_ack)&&((~o_wb_stb)||(i_wb_stall))) //&&(o_wb_cyc)
|
| 251 |
|
|
r_acks_waiting <= r_acks_waiting + {(LGCACHELEN+1){1'b1}}; // - 1;
|
| 252 |
|
|
|
| 253 |
|
|
always @(posedge i_clk)
|
| 254 |
|
|
if ((o_wb_cyc)&&(i_wb_ack))
|
| 255 |
|
|
cache[r_nvalid[(LGCACHELEN-1):0]+w_cache_offset]
|
| 256 |
|
|
<= i_wb_data;
|
| 257 |
|
|
|
| 258 |
|
|
initial r_addr_set = 1'b0;
|
| 259 |
|
|
always @(posedge i_clk)
|
| 260 |
|
|
if ((i_rst)||(i_new_pc))
|
| 261 |
|
|
r_addr_set <= 1'b1;
|
| 262 |
|
|
else if (i_clear_cache)
|
| 263 |
|
|
r_addr_set <= 1'b0;
|
| 264 |
|
|
|
| 265 |
|
|
// Now, read from the cache
|
| 266 |
|
|
wire w_cv; // Cache valid, address is in the cache
|
| 267 |
|
|
reg r_cv;
|
| 268 |
|
|
assign w_cv = ((r_nvalid != 0)&&(r_addr>=r_cache_base)
|
| 269 |
|
|
&&(r_addr-r_cache_base < bus_nvalid));
|
| 270 |
|
|
always @(posedge i_clk)
|
| 271 |
|
|
r_cv <= (~i_new_pc)&&((w_cv)||((~i_stall_n)&&(r_cv)));
|
| 272 |
|
|
assign o_v = (r_cv)&&(~i_new_pc);
|
| 273 |
|
|
|
| 274 |
|
|
always @(posedge i_clk)
|
| 275 |
|
|
if (i_new_pc)
|
| 276 |
|
|
r_addr <= i_pc;
|
| 277 |
|
|
else if ( ((i_stall_n)&&(w_cv)) || ((~i_stall_n)&&(w_cv)&&(r_addr == o_pc)) )
|
| 278 |
|
|
r_addr <= r_addr + {{(AW-1){1'b0}},1'b1};
|
| 279 |
|
|
|
| 280 |
|
|
wire [(LGCACHELEN-1):0] c_rdaddr, c_cache_base;
|
| 281 |
|
|
assign c_cache_base = r_cache_base[(LGCACHELEN-1):0];
|
| 282 |
|
|
assign c_rdaddr = r_addr[(LGCACHELEN-1):0]-c_cache_base+w_cache_offset;
|
| 283 |
|
|
always @(posedge i_clk)
|
| 284 |
|
|
if ((~o_v)||((i_stall_n)&&(o_v)))
|
| 285 |
|
|
o_i <= cache[c_rdaddr];
|
| 286 |
|
|
always @(posedge i_clk)
|
| 287 |
|
|
if ((~o_v)||((i_stall_n)&&(o_v)))
|
| 288 |
|
|
o_pc <= r_addr;
|
| 289 |
|
|
|
| 290 |
|
|
reg [(AW-1):0] ill_address;
|
| 291 |
|
|
initial ill_address = 0;
|
| 292 |
|
|
always @(posedge i_clk)
|
| 293 |
|
|
if ((o_wb_cyc)&&(i_wb_err))
|
| 294 |
|
|
ill_address <= o_wb_addr - {{(AW-LGCACHELEN-1){1'b0}}, r_acks_waiting};
|
| 295 |
|
|
|
| 296 |
|
|
assign o_illegal = (o_pc == ill_address)&&(~i_rst)&&(~i_new_pc)&&(~i_clear_cache);
|
| 297 |
|
|
|
| 298 |
|
|
|
| 299 |
|
|
endmodule
|
