| 1 |
2 |
smjoshua |
//////////////////////////////////////////////////////////////////
|
| 2 |
|
|
// //
|
| 3 |
|
|
// OoOPs Core Register Map Table module //
|
| 4 |
|
|
// //
|
| 5 |
|
|
// This file is part of the OoOPs project //
|
| 6 |
|
|
// http://www.opencores.org/project,oops //
|
| 7 |
|
|
// //
|
| 8 |
|
|
// Description: //
|
| 9 |
|
|
// The Map Table is responsible for maintaining the mapping //
|
| 10 |
|
|
// from architectural->physical registers. This block //
|
| 11 |
|
|
// consists of a free list for allocating new physical //
|
| 12 |
|
|
// registers and also the tables for mapping source operands. //
|
| 13 |
|
|
// //
|
| 14 |
|
|
// To avoid excessive flop usage for the map tables, block rams//
|
| 15 |
|
|
// will be used instead. //
|
| 16 |
|
|
// //
|
| 17 |
|
|
// Author(s): //
|
| 18 |
|
|
// - Joshua Smith, smjoshua@umich.edu //
|
| 19 |
|
|
// //
|
| 20 |
|
|
//////////////////////////////////////////////////////////////////
|
| 21 |
|
|
// //
|
| 22 |
|
|
// Copyright (C) 2012 Authors and OPENCORES.ORG //
|
| 23 |
|
|
// //
|
| 24 |
|
|
// This source file may be used and distributed without //
|
| 25 |
|
|
// restriction provided that this copyright statement is not //
|
| 26 |
|
|
// removed from the file and that any derivative work contains //
|
| 27 |
|
|
// the original copyright notice and the associated disclaimer. //
|
| 28 |
|
|
// //
|
| 29 |
|
|
// This source file is free software; you can redistribute it //
|
| 30 |
|
|
// and/or modify it under the terms of the GNU Lesser General //
|
| 31 |
|
|
// Public License as published by the Free Software Foundation; //
|
| 32 |
|
|
// either version 2.1 of the License, or (at your option) any //
|
| 33 |
|
|
// later version. //
|
| 34 |
|
|
// //
|
| 35 |
|
|
// This source is distributed in the hope that it will be //
|
| 36 |
|
|
// useful, but WITHOUT ANY WARRANTY; without even the implied //
|
| 37 |
|
|
// warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR //
|
| 38 |
|
|
// PURPOSE. See the GNU Lesser General Public License for more //
|
| 39 |
|
|
// details. //
|
| 40 |
|
|
// //
|
| 41 |
|
|
// You should have received a copy of the GNU Lesser General //
|
| 42 |
|
|
// Public License along with this source; if not, download it //
|
| 43 |
|
|
// from http://www.opencores.org/lgpl.shtml //
|
| 44 |
|
|
// //
|
| 45 |
|
|
//////////////////////////////////////////////////////////////////
|
| 46 |
|
|
`include "ooops_defs.v"
|
| 47 |
|
|
|
| 48 |
|
|
module map_table (
|
| 49 |
|
|
input wire clk,
|
| 50 |
|
|
input wire rst,
|
| 51 |
|
|
output wire map_table_init,
|
| 52 |
|
|
|
| 53 |
|
|
// Rename port
|
| 54 |
|
|
input wire ds1_valid,
|
| 55 |
|
|
input wire [`REG_IDX_SZ-1:0] ds1_src1_idx,
|
| 56 |
|
|
input wire [`REG_IDX_SZ-1:0] ds1_src2_idx,
|
| 57 |
|
|
input wire [`REG_IDX_SZ-1:0] ds1_dest_idx,
|
| 58 |
|
|
input wire ds1_dest_wr,
|
| 59 |
|
|
input wire ds1_type_br,
|
| 60 |
|
|
output wire [`TAG_SZ-1:0] ds2_src1_tag,
|
| 61 |
|
|
output wire [`TAG_SZ-1:0] ds2_src2_tag,
|
| 62 |
|
|
output wire ds2_src1_valid,
|
| 63 |
|
|
output wire ds2_src2_valid,
|
| 64 |
|
|
output wire [`TAG_SZ-1:0] ds2_dest_tag,
|
| 65 |
|
|
output wire [`TAG_SZ-1:0] ds2_dest_tag_old,
|
| 66 |
|
|
output wire [`FL_PTR_SZ-1:0] ds2_fl_head_ptr,
|
| 67 |
|
|
output wire [`CHKPT_PTR_SZ-1:0] ds2_chkpt_ptr,
|
| 68 |
|
|
|
| 69 |
|
|
// Writeback port
|
| 70 |
|
|
//input wire [`CDB_BUS_SZ-1:0] ex_cdb_bus,
|
| 71 |
|
|
|
| 72 |
|
|
// Retire and flush port
|
| 73 |
|
|
input wire rob_pipe_flush,
|
| 74 |
|
|
input wire rob_ds_ret_valid,
|
| 75 |
|
|
input wire rob_ds_ret_dest_write,
|
| 76 |
|
|
input wire [`CHKPT_PTR_SZ-1:0] rob_ds_chkpt_ptr,
|
| 77 |
|
|
input wire [`FL_PTR_SZ-1:0] rob_ds_fl_head_ptr,
|
| 78 |
|
|
input wire rob_ds_ret_chkpt_free,
|
| 79 |
|
|
input wire [`REG_IDX_SZ-1:0] rob_ds_ret_idx,
|
| 80 |
|
|
input wire [`TAG_SZ-1:0] rob_ds_ret_tag,
|
| 81 |
|
|
input wire [`TAG_SZ-1:0] rob_ds_ret_tag_old
|
| 82 |
|
|
);
|
| 83 |
|
|
|
| 84 |
|
|
// Internal wires and regs
|
| 85 |
|
|
wire [`TAG_SZ-1:0] ds1_dest_tag;
|
| 86 |
|
|
wire [`FL_PTR_SZ-1:0] ds1_fl_head_ptr;
|
| 87 |
|
|
wire [`ARCH_REGS-1:0] dfa_dirty_bit [`CHKPT_NUM-1:0];
|
| 88 |
|
|
wire [`ARCH_REGS-1:0] dfa_dirty_bit_in [`CHKPT_NUM-1:0];
|
| 89 |
|
|
wire [`CHKPT_NUM-1:0] dfa_dirty_bit_ld; // load is per checkpoint/column
|
| 90 |
|
|
wire [`CHKPT_NUM-1:0] dfa_dirty_bit_row [`ARCH_REGS-1:0];
|
| 91 |
|
|
wire [`CHKPT_PTR_SZ-1:0] ds1_src1_chkpt, ds1_src2_chkpt, ds1_dest_chkpt;
|
| 92 |
|
|
|
| 93 |
|
|
wire [`CHKPT_PTR_SZ-1:0] chkpt_head_ptr;
|
| 94 |
|
|
wire [`CHKPT_PTR_SZ-1:0] chkpt_head_ptr_p1;
|
| 95 |
|
|
wire [`CHKPT_PTR_SZ-1:0] chkpt_tail_ptr;
|
| 96 |
|
|
wire [`CHKPT_PTR_SZ-1:0] chkpt_tail_ptr_p1;
|
| 97 |
|
|
wire [`CHKPT_NUM-1:0] chkpt_valid_mask;
|
| 98 |
|
|
wire [`CHKPT_NUM-1:0] chkpt_valid_mask_in;
|
| 99 |
|
|
wire chkpt_allocate;
|
| 100 |
|
|
|
| 101 |
|
|
genvar g,k;
|
| 102 |
|
|
|
| 103 |
|
|
// Instantiate free list
|
| 104 |
|
|
free_list fl (
|
| 105 |
|
|
.clk(clk),
|
| 106 |
|
|
.rst(rst),
|
| 107 |
|
|
.ds1_dest_wr(ds1_dest_wr),
|
| 108 |
|
|
.rob_pipe_flush(rob_pipe_flush),
|
| 109 |
|
|
.rob_ds_fl_head_ptr(rob_ds_fl_head_ptr),
|
| 110 |
|
|
.rob_ds_ret_valid(rob_ds_ret_valid),
|
| 111 |
|
|
.rob_ds_ret_dest_write(rob_ds_ret_dest_write),
|
| 112 |
|
|
.rob_ds_ret_tag_old(rob_ds_ret_tag_old),
|
| 113 |
|
|
.ds1_dest_tag(ds1_dest_tag),
|
| 114 |
|
|
.ds1_fl_head_ptr(ds1_fl_head_ptr)
|
| 115 |
|
|
);
|
| 116 |
|
|
|
| 117 |
|
|
// Maintain the checkpoint head/tail pointers
|
| 118 |
|
|
// Operation:
|
| 119 |
|
|
// - Upon a pipe flush, restore both head and tail pointers to same pointer value from the ROB.
|
| 120 |
|
|
// - When a new checkpoint is allocated, advance head pointer by 1
|
| 121 |
|
|
// - When an instruction which allocated a checkpoint retires, advance tail pointer by 1.
|
| 122 |
|
|
assign chkpt_allocate = ds1_type_br & ~rob_pipe_flush;
|
| 123 |
|
|
wire [`CHKPT_PTR_SZ-1:0] chkpt_head_ptr_in = rob_pipe_flush ? rob_ds_chkpt_ptr : chkpt_head_ptr_p1;
|
| 124 |
|
|
wire [`CHKPT_PTR_SZ-1:0] chkpt_tail_ptr_in = rob_pipe_flush ? rob_ds_chkpt_ptr : chkpt_tail_ptr_p1;
|
| 125 |
|
|
wire chkpt_head_ptr_ld = rob_pipe_flush | chkpt_allocate;
|
| 126 |
|
|
wire chkpt_tail_ptr_ld = rob_pipe_flush | (rob_ds_ret_valid & rob_ds_ret_chkpt_free);
|
| 127 |
|
|
|
| 128 |
|
|
MDFFLR #(`CHKPT_PTR_SZ) chkpt_head_ptr_ff (clk, rst, chkpt_head_ptr_ld, `CHKPT_PTR_SZ'h0, chkpt_head_ptr_in, chkpt_head_ptr);
|
| 129 |
|
|
MDFFLR #(`CHKPT_PTR_SZ) chkpt_tail_ptr_ff (clk, rst, chkpt_tail_ptr_ld, `CHKPT_PTR_SZ'h0, chkpt_tail_ptr_in, chkpt_tail_ptr);
|
| 130 |
|
|
|
| 131 |
|
|
assign chkpt_head_ptr_p1 = (chkpt_head_ptr == `CHKPT_NUM-1) ? `CHKPT_PTR_SZ'h0 : chkpt_head_ptr + `CHKPT_PTR_SZ'h1;
|
| 132 |
|
|
assign chkpt_tail_ptr_p1 = (chkpt_tail_ptr == `CHKPT_NUM-1) ? `CHKPT_PTR_SZ'h0 : chkpt_tail_ptr + `CHKPT_PTR_SZ'h1;
|
| 133 |
|
|
|
| 134 |
|
|
// Keep a bit-vector mask of valid (allocated) checkpoints for the DFA search
|
| 135 |
|
|
// Initialize checkpoint 0 to valid, this will be the checkpoint used out of reset.
|
| 136 |
|
|
wire [`CHKPT_NUM-1:0] allocated_chkpt = (`CHKPT_NUM'h1 << chkpt_head_ptr_p1) & {`CHKPT_NUM{chkpt_allocate}};
|
| 137 |
|
|
wire [`CHKPT_NUM-1:0] freed_chkpt = (`CHKPT_NUM'h1 << chkpt_tail_ptr) & {`CHKPT_NUM{rob_ds_ret_chkpt_free}};
|
| 138 |
|
|
wire [`CHKPT_NUM-1:0] rob_ds_chkpt_vec = (`CHKPT_NUM'h1 << rob_ds_chkpt_ptr);
|
| 139 |
|
|
assign chkpt_valid_mask_in = rob_pipe_flush ? rob_ds_chkpt_vec : ((chkpt_valid_mask | allocated_chkpt) & ~freed_chkpt);
|
| 140 |
|
|
MDFFLR #(`CHKPT_NUM) chkpt_valid_mask_ff (clk, rst, chkpt_head_ptr_ld | chkpt_tail_ptr_ld, `CHKPT_NUM'h1, chkpt_valid_mask_in, chkpt_valid_mask);
|
| 141 |
|
|
|
| 142 |
|
|
/*
|
| 143 |
|
|
Handle the DFA (Dirty Flag Array) for determining which checkpoint contains the
|
| 144 |
|
|
most recent mapping for an architectural register. This is needed to setup the
|
| 145 |
|
|
SRAM address input for the RAT lookup.
|
| 146 |
|
|
|
| 147 |
|
|
Structure:
|
| 148 |
|
|
- Maintain a grid of bits (one row for each arch. reg, one column for each checkpoint).
|
| 149 |
|
|
- Head/Tail pointer keep track of most recently/least recently allocated valid checkpoints.
|
| 150 |
|
|
|
| 151 |
|
|
Operation:
|
| 152 |
|
|
- When a new checkpoint is allocated for a branch or speculation point, we advance
|
| 153 |
|
|
the head pointer and clear the entire DFA column for that checkpoint. For branches which
|
| 154 |
|
|
write a register, the write should update the old checkpoint, not the newly allocated one.
|
| 155 |
|
|
- When a register write operation comes through, update the row of the head checkpoint
|
| 156 |
|
|
corresponding to the destination architectural register index.
|
| 157 |
|
|
|
| 158 |
|
|
*/
|
| 159 |
|
|
wire [`CHKPT_NUM-1:0] dfa_column_clear = allocated_chkpt;
|
| 160 |
|
|
wire [`CHKPT_NUM-1:0] ds1_active_chkpt = (`CHKPT_NUM'h1 << chkpt_head_ptr);
|
| 161 |
|
|
assign dfa_dirty_bit_ld = dfa_column_clear | // Clear newly allocated checkpoint
|
| 162 |
|
|
(ds1_active_chkpt & {`CHKPT_NUM{ds1_dest_wr}}); // Update current checkpoint
|
| 163 |
|
|
|
| 164 |
|
|
wire [`ARCH_REGS-1:0] ds1_dest_idx_vec = (1 << ds1_dest_idx);
|
| 165 |
|
|
generate
|
| 166 |
|
|
for (g=0; g<`CHKPT_NUM; g=g+1) begin : dfa_gen
|
| 167 |
|
|
for (k=0; k<`ARCH_REGS; k=k+1) begin : dfa_dirty_bit_gen
|
| 168 |
|
|
assign dfa_dirty_bit_in[g][k] = ~dfa_column_clear[g] & ((ds1_dest_idx_vec[k] & ds1_dest_wr) ? 1'b1 : dfa_dirty_bit[g][k]);
|
| 169 |
|
|
|
| 170 |
|
|
MDFFLR #(1) dfa_dirty_bit_ff (clk, rst, dfa_dirty_bit_ld[g], 1'b0, dfa_dirty_bit_in[g][k], dfa_dirty_bit[g][k]);
|
| 171 |
|
|
|
| 172 |
|
|
// generate a "row" version as well
|
| 173 |
|
|
assign dfa_dirty_bit_row[k][g] = dfa_dirty_bit[g][k] & chkpt_valid_mask[g];
|
| 174 |
|
|
end
|
| 175 |
|
|
end
|
| 176 |
|
|
endgenerate
|
| 177 |
|
|
|
| 178 |
|
|
// Determine which checkpoint contains the most recent mapping for each source and the destination
|
| 179 |
|
|
// TODO: For now, assume 4 checkpoints. Find a nice way to make this general.
|
| 180 |
|
|
assign ds1_src1_chkpt = (chkpt_head_ptr == 2'h0) ? (dfa_dirty_bit_row[ds1_src1_idx][0] ? 2'h0 :
|
| 181 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][1] ? 2'h1 :
|
| 182 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][2] ? 2'h2 : 2'h3) :
|
| 183 |
|
|
(chkpt_head_ptr == 2'h1) ? (dfa_dirty_bit_row[ds1_src1_idx][1] ? 2'h1 :
|
| 184 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][2] ? 2'h2 :
|
| 185 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][3] ? 2'h3 : 2'h0) :
|
| 186 |
|
|
(chkpt_head_ptr == 2'h2) ? (dfa_dirty_bit_row[ds1_src1_idx][2] ? 2'h2 :
|
| 187 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][3] ? 2'h3 :
|
| 188 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][0] ? 2'h0 : 2'h1) :
|
| 189 |
|
|
(dfa_dirty_bit_row[ds1_src1_idx][3] ? 2'h3 :
|
| 190 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][0] ? 2'h0 :
|
| 191 |
|
|
dfa_dirty_bit_row[ds1_src1_idx][1] ? 2'h1 : 2'h2);
|
| 192 |
|
|
assign ds1_src2_chkpt = (chkpt_head_ptr == 2'h0) ? (dfa_dirty_bit_row[ds1_src2_idx][0] ? 2'h0 :
|
| 193 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][1] ? 2'h1 :
|
| 194 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][2] ? 2'h2 : 2'h3) :
|
| 195 |
|
|
(chkpt_head_ptr == 2'h1) ? (dfa_dirty_bit_row[ds1_src2_idx][1] ? 2'h1 :
|
| 196 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][2] ? 2'h2 :
|
| 197 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][3] ? 2'h3 : 2'h0) :
|
| 198 |
|
|
(chkpt_head_ptr == 2'h2) ? (dfa_dirty_bit_row[ds1_src2_idx][2] ? 2'h2 :
|
| 199 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][3] ? 2'h3 :
|
| 200 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][0] ? 2'h0 : 2'h1) :
|
| 201 |
|
|
(dfa_dirty_bit_row[ds1_src2_idx][3] ? 2'h3 :
|
| 202 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][0] ? 2'h0 :
|
| 203 |
|
|
dfa_dirty_bit_row[ds1_src2_idx][1] ? 2'h1 : 2'h2);
|
| 204 |
|
|
assign ds1_dest_chkpt = (chkpt_head_ptr == 2'h0) ? (dfa_dirty_bit_row[ds1_dest_idx][0] ? 2'h0 :
|
| 205 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][1] ? 2'h1 :
|
| 206 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][2] ? 2'h2 : 2'h3) :
|
| 207 |
|
|
(chkpt_head_ptr == 2'h1) ? (dfa_dirty_bit_row[ds1_dest_idx][1] ? 2'h1 :
|
| 208 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][2] ? 2'h2 :
|
| 209 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][3] ? 2'h3 : 2'h0) :
|
| 210 |
|
|
(chkpt_head_ptr == 2'h2) ? (dfa_dirty_bit_row[ds1_dest_idx][2] ? 2'h2 :
|
| 211 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][3] ? 2'h3 :
|
| 212 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][0] ? 2'h0 : 2'h1) :
|
| 213 |
|
|
(dfa_dirty_bit_row[ds1_dest_idx][3] ? 2'h3 :
|
| 214 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][0] ? 2'h0 :
|
| 215 |
|
|
dfa_dirty_bit_row[ds1_dest_idx][1] ? 2'h1 : 2'h2);
|
| 216 |
|
|
|
| 217 |
|
|
// If no dirty bit set for any of the valid checkpoints, then committed copy must have latest mapping
|
| 218 |
|
|
wire ds1_src1_use_rrat = ~(|dfa_dirty_bit_row[ds1_src1_idx]);
|
| 219 |
|
|
wire ds1_src2_use_rrat = ~(|dfa_dirty_bit_row[ds1_src2_idx]);
|
| 220 |
|
|
wire ds1_dest_use_rrat = ~(|dfa_dirty_bit_row[ds1_dest_idx]);
|
| 221 |
|
|
|
| 222 |
|
|
// Generate the RAT SRAM read/write addresses and controls
|
| 223 |
|
|
// Note: since tables are SRAM-based, we need to initialize the RRAT so that
|
| 224 |
|
|
// registers are mapped correctly out of reset
|
| 225 |
|
|
wire [`REG_IDX_SZ-1:0] map_table_init_ctr, map_table_init_ctr_in;
|
| 226 |
|
|
wire map_table_init_in = map_table_init & (map_table_init_ctr != `ARCH_REGS);
|
| 227 |
|
|
MDFFR #(1) map_table_init_ff (clk, rst, 1'b1, map_table_init_in, map_table_init);
|
| 228 |
|
|
|
| 229 |
|
|
assign map_table_init_ctr_in = map_table_init_ctr + `REG_IDX_SZ'h1;
|
| 230 |
|
|
MDFFLR #(`REG_IDX_SZ) map_table_init_ctr_ff (clk, rst, map_table_init, `REG_IDX_SZ'h0, map_table_init_ctr_in, map_table_init_ctr);
|
| 231 |
|
|
|
| 232 |
|
|
wire [`REG_IDX_SZ+`CHKPT_PTR_SZ-1:0] ds1_rat_src1_rd_addr = {ds1_src1_idx,ds1_src1_chkpt};
|
| 233 |
|
|
wire [`REG_IDX_SZ+`CHKPT_PTR_SZ-1:0] ds1_rat_src2_rd_addr = {ds1_src2_idx,ds1_src2_chkpt};
|
| 234 |
|
|
wire [`REG_IDX_SZ+`CHKPT_PTR_SZ-1:0] ds1_rat_dest_rd_addr = {ds1_dest_idx,ds1_dest_chkpt};
|
| 235 |
|
|
|
| 236 |
|
|
// Writes need to come from DS2 stage in case we read and write the same arch. register
|
| 237 |
|
|
wire ds2_rat_wren, ds2_rat_wren_in;
|
| 238 |
|
|
wire [`REG_IDX_SZ+`CHKPT_PTR_SZ-1:0] ds2_rat_wr_addr, ds2_rat_wr_addr_in;
|
| 239 |
|
|
wire [`TAG_SZ-1:0] ds2_rat_wr_data;
|
| 240 |
|
|
|
| 241 |
|
|
assign ds2_rat_wren_in = ds1_dest_wr;
|
| 242 |
|
|
assign ds2_rat_wr_addr_in = {ds1_dest_idx,chkpt_head_ptr};
|
| 243 |
|
|
MDFFR #(1) ds2_rat_wren_ff (clk, rst, 1'b0, ds2_rat_wren_in, ds2_rat_wren);
|
| 244 |
|
|
MDFFR #(`REG_IDX_SZ+`CHKPT_PTR_SZ) ds2_rat_wr_addr_ff (clk, rst, 1'b0, ds2_rat_wr_addr_in, ds2_rat_wr_addr);
|
| 245 |
|
|
assign ds2_rat_wr_data = ds2_dest_tag;
|
| 246 |
|
|
|
| 247 |
|
|
wire [`TAG_SZ-1:0] ds2_rat_src1_rd_data, ds2_rrat_src1_rd_data;
|
| 248 |
|
|
wire [`TAG_SZ-1:0] ds2_rat_src2_rd_data, ds2_rrat_src2_rd_data;
|
| 249 |
|
|
wire [`TAG_SZ-1:0] ds2_rat_dest_rd_data, ds2_rrat_dest_rd_data;
|
| 250 |
|
|
|
| 251 |
|
|
wire [`REG_IDX_SZ-1:0] ds_rrat_wr_addr = map_table_init ? map_table_init_ctr : rob_ds_ret_idx;
|
| 252 |
|
|
wire [`TAG_SZ-1:0] ds_rrat_wr_data = map_table_init ? map_table_init_ctr : rob_ds_ret_tag;
|
| 253 |
|
|
wire ds_rrat_wren = map_table_init | rob_ds_ret_valid & rob_ds_ret_dest_write;
|
| 254 |
|
|
|
| 255 |
|
|
// Instantiate RAT SRAM blocks
|
| 256 |
|
|
// Note that we need 3 copies for the required 3 read ports (2 source operand tag reads, 1 previous dest tag read)
|
| 257 |
|
|
// Read copy 1
|
| 258 |
|
|
dp_sram #(.DW(`TAG_SZ), .IW(`REG_IDX_SZ+`CHKPT_PTR_SZ)) rat0 (
|
| 259 |
|
|
.clk(clk),
|
| 260 |
|
|
.a_addr(ds1_rat_src1_rd_addr), // Read port
|
| 261 |
|
|
.a_dout(ds2_rat_src1_rd_data),
|
| 262 |
|
|
|
| 263 |
|
|
.b_addr(ds2_rat_wr_addr), // Write port
|
| 264 |
|
|
.b_wren(ds2_rat_wren),
|
| 265 |
|
|
.b_din(ds2_rat_wr_data)
|
| 266 |
|
|
);
|
| 267 |
|
|
|
| 268 |
|
|
// Read copy 2
|
| 269 |
|
|
dp_sram #(.DW(`TAG_SZ), .IW(`REG_IDX_SZ+`CHKPT_PTR_SZ)) rat1 (
|
| 270 |
|
|
.clk(clk),
|
| 271 |
|
|
.a_addr(ds1_rat_src2_rd_addr), // Read port
|
| 272 |
|
|
.a_dout(ds2_rat_src2_rd_data),
|
| 273 |
|
|
|
| 274 |
|
|
.b_addr(ds2_rat_wr_addr), // Write port
|
| 275 |
|
|
.b_wren(ds2_rat_wren),
|
| 276 |
|
|
.b_din(ds2_rat_wr_data)
|
| 277 |
|
|
);
|
| 278 |
|
|
|
| 279 |
|
|
// Write copy 1
|
| 280 |
|
|
dp_sram #(.DW(`TAG_SZ), .IW(`REG_IDX_SZ+`CHKPT_PTR_SZ)) rat2 (
|
| 281 |
|
|
.clk(clk),
|
| 282 |
|
|
.a_addr(ds1_rat_dest_rd_addr), // Read port
|
| 283 |
|
|
.a_dout(ds2_rat_dest_rd_data),
|
| 284 |
|
|
|
| 285 |
|
|
.b_addr(ds2_rat_wr_addr), // Write port
|
| 286 |
|
|
.b_wren(ds2_rat_wren),
|
| 287 |
|
|
.b_din(ds2_rat_wr_data)
|
| 288 |
|
|
);
|
| 289 |
|
|
|
| 290 |
|
|
// Instantiate tables for the committed RAT copies
|
| 291 |
|
|
dp_sram #(.DW(`TAG_SZ), .IW(`REG_IDX_SZ)) rrat0 (
|
| 292 |
|
|
.clk(clk),
|
| 293 |
|
|
.a_addr(ds1_src1_idx), // Read port
|
| 294 |
|
|
.a_dout(ds2_rrat_src1_rd_data),
|
| 295 |
|
|
|
| 296 |
|
|
.b_addr(ds_rrat_wr_addr), // Write port (controlled by retire)
|
| 297 |
|
|
.b_wren(ds_rrat_wren),
|
| 298 |
|
|
.b_din(ds_rrat_wr_data)
|
| 299 |
|
|
);
|
| 300 |
|
|
dp_sram #(.DW(`TAG_SZ), .IW(`REG_IDX_SZ)) rrat1 (
|
| 301 |
|
|
.clk(clk),
|
| 302 |
|
|
.a_addr(ds1_src2_idx), // Read port
|
| 303 |
|
|
.a_dout(ds2_rrat_src2_rd_data),
|
| 304 |
|
|
|
| 305 |
|
|
.b_addr(ds_rrat_wr_addr), // Write port (controlled by retire)
|
| 306 |
|
|
.b_wren(ds_rrat_wren),
|
| 307 |
|
|
.b_din(ds_rrat_wr_data)
|
| 308 |
|
|
);
|
| 309 |
|
|
dp_sram #(.DW(`TAG_SZ), .IW(`REG_IDX_SZ)) rrat2 (
|
| 310 |
|
|
.clk(clk),
|
| 311 |
|
|
.a_addr(ds1_dest_idx), // Read port
|
| 312 |
|
|
.a_dout(ds2_rrat_dest_rd_data),
|
| 313 |
|
|
|
| 314 |
|
|
.b_addr(ds_rrat_wr_addr), // Write port (controlled by retire)
|
| 315 |
|
|
.b_wren(ds_rrat_wren),
|
| 316 |
|
|
.b_din(ds_rrat_wr_data)
|
| 317 |
|
|
);
|
| 318 |
|
|
|
| 319 |
|
|
// Since writes to map tables occur in DS2 stage, need to detect forwarding from previous instructions
|
| 320 |
|
|
wire ds1_src1_wr_fwd = (ds1_src1_idx == ds2_rat_wr_addr[`REG_IDX_SZ+`CHKPT_PTR_SZ-1:`CHKPT_PTR_SZ]) & ds2_rat_wren;
|
| 321 |
|
|
wire ds1_src2_wr_fwd = (ds1_src2_idx == ds2_rat_wr_addr[`REG_IDX_SZ+`CHKPT_PTR_SZ-1:`CHKPT_PTR_SZ]) & ds2_rat_wren;
|
| 322 |
|
|
wire ds1_dest_wr_fwd = (ds1_dest_idx == ds2_rat_wr_addr[`REG_IDX_SZ+`CHKPT_PTR_SZ-1:`CHKPT_PTR_SZ]) & ds2_rat_wren;
|
| 323 |
|
|
wire ds2_src1_wr_fwd, ds2_src2_wr_fwd, ds2_dest_wr_fwd;
|
| 324 |
|
|
MDFFR #(1) ds2_src1_wr_fwd_ff (clk, rst, 1'b0, ds1_src1_wr_fwd, ds2_src1_wr_fwd);
|
| 325 |
|
|
MDFFR #(1) ds2_src2_wr_fwd_ff (clk, rst, 1'b0, ds1_src2_wr_fwd, ds2_src2_wr_fwd);
|
| 326 |
|
|
MDFFR #(1) ds2_dest_wr_fwd_ff (clk, rst, 1'b0, ds1_dest_wr_fwd, ds2_dest_wr_fwd);
|
| 327 |
|
|
|
| 328 |
|
|
wire [`TAG_SZ-1:0] r_ds2_rat_wr_data;
|
| 329 |
|
|
wire r_ds2_wr_data_ld = ds2_rat_wren & (ds1_src1_wr_fwd | ds1_src2_wr_fwd | ds1_dest_wr_fwd);
|
| 330 |
|
|
MDFFL #(`TAG_SZ) r_ds2_rat_wr_data_ff (clk, r_ds2_wr_data_ld, ds2_rat_wr_data, r_ds2_rat_wr_data);
|
| 331 |
|
|
|
| 332 |
|
|
// Generate DS2 stage outputs
|
| 333 |
|
|
// Mux between RRAT and RAT outputs
|
| 334 |
|
|
MDFFL #(`CHKPT_PTR_SZ) ds2_chkpt_ptr_ff (clk, ds1_valid, chkpt_head_ptr, ds2_chkpt_ptr);
|
| 335 |
|
|
MDFFL #(`FL_PTR_SZ) ds2_fl_head_ptr_ff (clk, ds1_valid, ds1_fl_head_ptr, ds2_fl_head_ptr);
|
| 336 |
|
|
|
| 337 |
|
|
wire ds2_src1_use_rrat, ds2_src2_use_rrat, ds2_dest_use_rrat;
|
| 338 |
|
|
MDFFLR #(1) ds2_src1_use_rrat_ff (clk, rst, ds1_valid, 1'b0, ds1_src1_use_rrat, ds2_src1_use_rrat);
|
| 339 |
|
|
MDFFLR #(1) ds2_src2_use_rrat_ff (clk, rst, ds1_valid, 1'b0, ds1_src2_use_rrat, ds2_src2_use_rrat);
|
| 340 |
|
|
MDFFLR #(1) ds2_dest_use_rrat_ff (clk, rst, ds1_valid, 1'b0, ds1_dest_use_rrat, ds2_dest_use_rrat);
|
| 341 |
|
|
MDFFL #(`TAG_SZ) ds2_dest_tag_ff (clk, ds1_valid, ds1_dest_tag, ds2_dest_tag);
|
| 342 |
|
|
|
| 343 |
|
|
assign ds2_src1_tag = ds2_src1_wr_fwd ? r_ds2_rat_wr_data : ds2_src1_use_rrat ? ds2_rrat_src1_rd_data : ds2_rat_src1_rd_data;
|
| 344 |
|
|
assign ds2_src2_tag = ds2_src2_wr_fwd ? r_ds2_rat_wr_data : ds2_src2_use_rrat ? ds2_rrat_src2_rd_data : ds2_rat_src2_rd_data;
|
| 345 |
|
|
assign ds2_dest_tag_old = ds2_dest_wr_fwd ? r_ds2_rat_wr_data : ds2_dest_use_rrat ? ds2_rrat_dest_rd_data : ds2_rat_dest_rd_data;
|
| 346 |
|
|
|
| 347 |
|
|
|
| 348 |
|
|
|
| 349 |
|
|
|
| 350 |
|
|
endmodule
|