mirror of
https://github.com/lowRISC/ibex.git
synced 2025-04-20 11:57:12 -04:00
1400 lines
51 KiB
Systemverilog
1400 lines
51 KiB
Systemverilog
// Copyright lowRISC contributors.
|
|
// Copyright 2018 ETH Zurich and University of Bologna, see also CREDITS.md.
|
|
// Licensed under the Apache License, Version 2.0, see LICENSE for details.
|
|
// SPDX-License-Identifier: Apache-2.0
|
|
|
|
/**
|
|
* Arithmetic logic unit
|
|
*/
|
|
module ibex_alu #(
|
|
parameter ibex_pkg::rv32b_e RV32B = ibex_pkg::RV32BNone
|
|
) (
|
|
input ibex_pkg::alu_op_e operator_i,
|
|
input logic [31:0] operand_a_i,
|
|
input logic [31:0] operand_b_i,
|
|
|
|
input logic instr_first_cycle_i,
|
|
|
|
input logic [32:0] multdiv_operand_a_i,
|
|
input logic [32:0] multdiv_operand_b_i,
|
|
|
|
input logic multdiv_sel_i,
|
|
|
|
input logic [31:0] imd_val_q_i[2],
|
|
output logic [31:0] imd_val_d_o[2],
|
|
output logic [1:0] imd_val_we_o,
|
|
|
|
output logic [31:0] adder_result_o,
|
|
output logic [33:0] adder_result_ext_o,
|
|
|
|
output logic [31:0] result_o,
|
|
output logic comparison_result_o,
|
|
output logic is_equal_result_o
|
|
);
|
|
import ibex_pkg::*;
|
|
|
|
logic [31:0] operand_a_rev;
|
|
logic [32:0] operand_b_neg;
|
|
|
|
// bit reverse operand_a for left shifts and bit counting
|
|
for (genvar k = 0; k < 32; k++) begin : gen_rev_operand_a
|
|
assign operand_a_rev[k] = operand_a_i[31-k];
|
|
end
|
|
|
|
///////////
|
|
// Adder //
|
|
///////////
|
|
|
|
logic adder_op_a_shift1;
|
|
logic adder_op_a_shift2;
|
|
logic adder_op_a_shift3;
|
|
logic adder_op_b_negate;
|
|
logic [32:0] adder_in_a, adder_in_b;
|
|
logic [31:0] adder_result;
|
|
|
|
always_comb begin
|
|
adder_op_a_shift1 = 1'b0;
|
|
adder_op_a_shift2 = 1'b0;
|
|
adder_op_a_shift3 = 1'b0;
|
|
adder_op_b_negate = 1'b0;
|
|
unique case (operator_i)
|
|
// Adder OPs
|
|
ALU_SUB,
|
|
|
|
// Comparator OPs
|
|
ALU_EQ, ALU_NE,
|
|
ALU_GE, ALU_GEU,
|
|
ALU_LT, ALU_LTU,
|
|
ALU_SLT, ALU_SLTU,
|
|
|
|
// MinMax OPs (RV32B Ops)
|
|
ALU_MIN, ALU_MINU,
|
|
ALU_MAX, ALU_MAXU: adder_op_b_negate = 1'b1;
|
|
|
|
// Address Calculation OPs (RV32B Ops)
|
|
ALU_SH1ADD: if (RV32B != RV32BNone) adder_op_a_shift1 = 1'b1;
|
|
ALU_SH2ADD: if (RV32B != RV32BNone) adder_op_a_shift2 = 1'b1;
|
|
ALU_SH3ADD: if (RV32B != RV32BNone) adder_op_a_shift3 = 1'b1;
|
|
|
|
default:;
|
|
endcase
|
|
end
|
|
|
|
// prepare operand a
|
|
always_comb begin
|
|
unique case (1'b1)
|
|
multdiv_sel_i: adder_in_a = multdiv_operand_a_i;
|
|
adder_op_a_shift1: adder_in_a = {operand_a_i[30:0],2'b01};
|
|
adder_op_a_shift2: adder_in_a = {operand_a_i[29:0],3'b001};
|
|
adder_op_a_shift3: adder_in_a = {operand_a_i[28:0],4'b0001};
|
|
default: adder_in_a = {operand_a_i,1'b1};
|
|
endcase
|
|
end
|
|
|
|
// prepare operand b
|
|
assign operand_b_neg = {operand_b_i,1'b0} ^ {33{1'b1}};
|
|
always_comb begin
|
|
unique case (1'b1)
|
|
multdiv_sel_i: adder_in_b = multdiv_operand_b_i;
|
|
adder_op_b_negate: adder_in_b = operand_b_neg;
|
|
default: adder_in_b = {operand_b_i, 1'b0};
|
|
endcase
|
|
end
|
|
|
|
// actual adder
|
|
assign adder_result_ext_o = $unsigned(adder_in_a) + $unsigned(adder_in_b);
|
|
|
|
assign adder_result = adder_result_ext_o[32:1];
|
|
|
|
assign adder_result_o = adder_result;
|
|
|
|
////////////////
|
|
// Comparison //
|
|
////////////////
|
|
|
|
logic is_equal;
|
|
logic is_greater_equal; // handles both signed and unsigned forms
|
|
logic cmp_signed;
|
|
|
|
always_comb begin
|
|
unique case (operator_i)
|
|
ALU_GE,
|
|
ALU_LT,
|
|
ALU_SLT,
|
|
// RV32B only
|
|
ALU_MIN,
|
|
ALU_MAX: cmp_signed = 1'b1;
|
|
|
|
default: cmp_signed = 1'b0;
|
|
endcase
|
|
end
|
|
|
|
assign is_equal = (adder_result == 32'b0);
|
|
assign is_equal_result_o = is_equal;
|
|
|
|
// Is greater equal
|
|
always_comb begin
|
|
if ((operand_a_i[31] ^ operand_b_i[31]) == 1'b0) begin
|
|
is_greater_equal = (adder_result[31] == 1'b0);
|
|
end else begin
|
|
is_greater_equal = operand_a_i[31] ^ (cmp_signed);
|
|
end
|
|
end
|
|
|
|
// GTE unsigned:
|
|
// (a[31] == 1 && b[31] == 1) => adder_result[31] == 0
|
|
// (a[31] == 0 && b[31] == 0) => adder_result[31] == 0
|
|
// (a[31] == 1 && b[31] == 0) => 1
|
|
// (a[31] == 0 && b[31] == 1) => 0
|
|
|
|
// GTE signed:
|
|
// (a[31] == 1 && b[31] == 1) => adder_result[31] == 0
|
|
// (a[31] == 0 && b[31] == 0) => adder_result[31] == 0
|
|
// (a[31] == 1 && b[31] == 0) => 0
|
|
// (a[31] == 0 && b[31] == 1) => 1
|
|
|
|
// generate comparison result
|
|
logic cmp_result;
|
|
|
|
always_comb begin
|
|
unique case (operator_i)
|
|
ALU_EQ: cmp_result = is_equal;
|
|
ALU_NE: cmp_result = ~is_equal;
|
|
ALU_GE, ALU_GEU,
|
|
ALU_MAX, ALU_MAXU: cmp_result = is_greater_equal; // RV32B only
|
|
ALU_LT, ALU_LTU,
|
|
ALU_MIN, ALU_MINU, //RV32B only
|
|
ALU_SLT, ALU_SLTU: cmp_result = ~is_greater_equal;
|
|
|
|
default: cmp_result = is_equal;
|
|
endcase
|
|
end
|
|
|
|
assign comparison_result_o = cmp_result;
|
|
|
|
///////////
|
|
// Shift //
|
|
///////////
|
|
|
|
// The shifter structure consists of a 33-bit shifter: 32-bit operand + 1 bit extension for
|
|
// arithmetic shifts and one-shift support.
|
|
// Rotations and funnel shifts are implemented as multi-cycle instructions.
|
|
// The shifter is also used for single-bit instructions and bit-field place as detailed below.
|
|
//
|
|
// Standard Shifts
|
|
// ===============
|
|
// For standard shift instructions, the direction of the shift is to the right by default. For
|
|
// left shifts, the signal shift_left signal is set. If so, the operand is initially reversed,
|
|
// shifted to the right by the specified amount and shifted back again. For arithmetic- and
|
|
// one-shifts the 33rd bit of the shifter operand can is set accordingly.
|
|
//
|
|
// Multicycle Shifts
|
|
// =================
|
|
//
|
|
// Rotation
|
|
// --------
|
|
// For rotations, the operand signals operand_a_i and operand_b_i are kept constant to rs1 and
|
|
// rs2 respectively.
|
|
//
|
|
// Rotation pseudocode:
|
|
// shift_amt = rs2 & 31;
|
|
// multicycle_result = (rs1 >> shift_amt) | (rs1 << (32 - shift_amt));
|
|
// ^-- cycle 0 -----^ ^-- cycle 1 --------------^
|
|
//
|
|
// Funnel Shifts
|
|
// -------------
|
|
// For funnel shifs, operand_a_i is tied to rs1 in the first cycle and rs3 in the
|
|
// second cycle. operand_b_i is always tied to rs2. The order of applying the shift amount or
|
|
// its complement is determined by bit [5] of shift_amt.
|
|
//
|
|
// Funnel shift Pseudocode: (fsl)
|
|
// shift_amt = rs2 & 63;
|
|
// shift_amt_compl = 32 - shift_amt[4:0]
|
|
// if (shift_amt >=33):
|
|
// multicycle_result = (rs1 >> shift_amt_compl[4:0]) | (rs3 << shift_amt[4:0]);
|
|
// ^-- cycle 0 ----------------^ ^-- cycle 1 ------------^
|
|
// else if (shift_amt <= 31 && shift_amt > 0):
|
|
// multicycle_result = (rs1 << shift_amt[4:0]) | (rs3 >> shift_amt_compl[4:0]);
|
|
// ^-- cycle 0 ----------^ ^-- cycle 1 -------------------^
|
|
// For shift_amt == 0, 32, both shift_amt[4:0] and shift_amt_compl[4:0] == '0.
|
|
// these cases need to be handled separately outside the shifting structure:
|
|
// else if (shift_amt == 32):
|
|
// multicycle_result = rs3
|
|
// else if (shift_amt == 0):
|
|
// multicycle_result = rs1.
|
|
//
|
|
// Single-Bit Instructions
|
|
// =======================
|
|
// Single bit instructions operate on bit operand_b_i[4:0] of operand_a_i.
|
|
|
|
// The operations bset, bclr and binv are implemented by generation of a bit-mask using the
|
|
// shifter structure. This is done by left-shifting the operand 32'h1 by the required amount.
|
|
// The signal shift_sbmode multiplexes the shifter input and sets the signal shift_left.
|
|
// Further processing is taken care of by a separate structure.
|
|
//
|
|
// For bext, the bit defined by operand_b_i[4:0] is to be returned. This is done by simply
|
|
// shifting operand_a_i to the right by the required amount and returning bit [0] of the result.
|
|
//
|
|
// Bit-Field Place
|
|
// ===============
|
|
// The shifter structure is shared to compute bfp_mask << bfp_off.
|
|
|
|
logic shift_left;
|
|
logic shift_ones;
|
|
logic shift_arith;
|
|
logic shift_funnel;
|
|
logic shift_sbmode;
|
|
logic [5:0] shift_amt;
|
|
logic [5:0] shift_amt_compl; // complementary shift amount (32 - shift_amt)
|
|
|
|
logic [31:0] shift_operand;
|
|
logic signed [32:0] shift_result_ext_signed;
|
|
logic [32:0] shift_result_ext;
|
|
logic unused_shift_result_ext;
|
|
logic [31:0] shift_result;
|
|
logic [31:0] shift_result_rev;
|
|
|
|
// zbf
|
|
logic bfp_op;
|
|
logic [4:0] bfp_len;
|
|
logic [4:0] bfp_off;
|
|
logic [31:0] bfp_mask;
|
|
logic [31:0] bfp_mask_rev;
|
|
logic [31:0] bfp_result;
|
|
|
|
// bfp: shares the shifter structure to compute bfp_mask << bfp_off
|
|
assign bfp_op = (RV32B != RV32BNone) ? (operator_i == ALU_BFP) : 1'b0;
|
|
assign bfp_len = {~(|operand_b_i[27:24]), operand_b_i[27:24]}; // len = 0 encodes for len = 16
|
|
assign bfp_off = operand_b_i[20:16];
|
|
assign bfp_mask = (RV32B != RV32BNone) ? ~(32'hffff_ffff << bfp_len) : '0;
|
|
for (genvar i = 0; i < 32; i++) begin : gen_rev_bfp_mask
|
|
assign bfp_mask_rev[i] = bfp_mask[31-i];
|
|
end
|
|
|
|
assign bfp_result =(RV32B != RV32BNone) ?
|
|
(~shift_result & operand_a_i) | ((operand_b_i & bfp_mask) << bfp_off) : '0;
|
|
|
|
// bit shift_amt[5]: word swap bit: only considered for FSL/FSR.
|
|
// if set, reverse operations in first and second cycle.
|
|
assign shift_amt[5] = operand_b_i[5] & shift_funnel;
|
|
assign shift_amt_compl = 32 - operand_b_i[4:0];
|
|
|
|
always_comb begin
|
|
if (bfp_op) begin
|
|
shift_amt[4:0] = bfp_off; // length field of bfp control word
|
|
end else begin
|
|
shift_amt[4:0] = instr_first_cycle_i ?
|
|
(operand_b_i[5] && shift_funnel ? shift_amt_compl[4:0] : operand_b_i[4:0]) :
|
|
(operand_b_i[5] && shift_funnel ? operand_b_i[4:0] : shift_amt_compl[4:0]);
|
|
end
|
|
end
|
|
|
|
// single-bit mode: shift
|
|
assign shift_sbmode = (RV32B != RV32BNone) ?
|
|
(operator_i == ALU_BSET) | (operator_i == ALU_BCLR) | (operator_i == ALU_BINV) : 1'b0;
|
|
|
|
// left shift if this is:
|
|
// * a standard left shift (slo, sll)
|
|
// * a rol in the first cycle
|
|
// * a ror in the second cycle
|
|
// * fsl: without word-swap bit: first cycle, else: second cycle
|
|
// * fsr: without word-swap bit: second cycle, else: first cycle
|
|
// * a single-bit instruction: bclr, bset, binv (excluding bext)
|
|
// * bfp: bfp_mask << bfp_off
|
|
always_comb begin
|
|
unique case (operator_i)
|
|
ALU_SLL: shift_left = 1'b1;
|
|
ALU_SLO: shift_left = (RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) ? 1'b1 : 1'b0;
|
|
ALU_BFP: shift_left = (RV32B != RV32BNone) ? 1'b1 : 1'b0;
|
|
ALU_ROL: shift_left = (RV32B != RV32BNone) ? instr_first_cycle_i : 0;
|
|
ALU_ROR: shift_left = (RV32B != RV32BNone) ? ~instr_first_cycle_i : 0;
|
|
ALU_FSL: shift_left = (RV32B != RV32BNone) ?
|
|
(shift_amt[5] ? ~instr_first_cycle_i : instr_first_cycle_i) : 1'b0;
|
|
ALU_FSR: shift_left = (RV32B != RV32BNone) ?
|
|
(shift_amt[5] ? instr_first_cycle_i : ~instr_first_cycle_i) : 1'b0;
|
|
default: shift_left = 1'b0;
|
|
endcase
|
|
if (shift_sbmode) begin
|
|
shift_left = 1'b1;
|
|
end
|
|
end
|
|
|
|
assign shift_arith = (operator_i == ALU_SRA);
|
|
assign shift_ones = (RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) ?
|
|
(operator_i == ALU_SLO) | (operator_i == ALU_SRO) : 1'b0;
|
|
assign shift_funnel = (RV32B != RV32BNone) ?
|
|
(operator_i == ALU_FSL) | (operator_i == ALU_FSR) : 1'b0;
|
|
|
|
// shifter structure.
|
|
always_comb begin
|
|
// select shifter input
|
|
// for bfp, sbmode and shift_left the corresponding bit-reversed input is chosen.
|
|
if (RV32B == RV32BNone) begin
|
|
shift_operand = shift_left ? operand_a_rev : operand_a_i;
|
|
end else begin
|
|
unique case (1'b1)
|
|
bfp_op: shift_operand = bfp_mask_rev;
|
|
shift_sbmode: shift_operand = 32'h8000_0000;
|
|
default: shift_operand = shift_left ? operand_a_rev : operand_a_i;
|
|
endcase
|
|
end
|
|
|
|
shift_result_ext_signed =
|
|
$signed({shift_ones | (shift_arith & shift_operand[31]), shift_operand}) >>> shift_amt[4:0];
|
|
shift_result_ext = $unsigned(shift_result_ext_signed);
|
|
|
|
shift_result = shift_result_ext[31:0];
|
|
unused_shift_result_ext = shift_result_ext[32];
|
|
|
|
for (int unsigned i = 0; i < 32; i++) begin
|
|
shift_result_rev[i] = shift_result[31-i];
|
|
end
|
|
|
|
shift_result = shift_left ? shift_result_rev : shift_result;
|
|
|
|
end
|
|
|
|
///////////////////
|
|
// Bitwise Logic //
|
|
///////////////////
|
|
|
|
logic bwlogic_or;
|
|
logic bwlogic_and;
|
|
logic [31:0] bwlogic_operand_b;
|
|
logic [31:0] bwlogic_or_result;
|
|
logic [31:0] bwlogic_and_result;
|
|
logic [31:0] bwlogic_xor_result;
|
|
logic [31:0] bwlogic_result;
|
|
|
|
logic bwlogic_op_b_negate;
|
|
|
|
always_comb begin
|
|
unique case (operator_i)
|
|
// Logic-with-negate OPs (RV32B Ops)
|
|
ALU_XNOR,
|
|
ALU_ORN,
|
|
ALU_ANDN: bwlogic_op_b_negate = (RV32B != RV32BNone) ? 1'b1 : 1'b0;
|
|
ALU_CMIX: bwlogic_op_b_negate = (RV32B != RV32BNone) ? ~instr_first_cycle_i : 1'b0;
|
|
default: bwlogic_op_b_negate = 1'b0;
|
|
endcase
|
|
end
|
|
|
|
assign bwlogic_operand_b = bwlogic_op_b_negate ? operand_b_neg[32:1] : operand_b_i;
|
|
|
|
assign bwlogic_or_result = operand_a_i | bwlogic_operand_b;
|
|
assign bwlogic_and_result = operand_a_i & bwlogic_operand_b;
|
|
assign bwlogic_xor_result = operand_a_i ^ bwlogic_operand_b;
|
|
|
|
assign bwlogic_or = (operator_i == ALU_OR) | (operator_i == ALU_ORN);
|
|
assign bwlogic_and = (operator_i == ALU_AND) | (operator_i == ALU_ANDN);
|
|
|
|
always_comb begin
|
|
unique case (1'b1)
|
|
bwlogic_or: bwlogic_result = bwlogic_or_result;
|
|
bwlogic_and: bwlogic_result = bwlogic_and_result;
|
|
default: bwlogic_result = bwlogic_xor_result;
|
|
endcase
|
|
end
|
|
|
|
logic [5:0] bitcnt_result;
|
|
logic [31:0] minmax_result;
|
|
logic [31:0] pack_result;
|
|
logic [31:0] sext_result;
|
|
logic [31:0] singlebit_result;
|
|
logic [31:0] rev_result;
|
|
logic [31:0] shuffle_result;
|
|
logic [31:0] xperm_result;
|
|
logic [31:0] butterfly_result;
|
|
logic [31:0] invbutterfly_result;
|
|
logic [31:0] clmul_result;
|
|
logic [31:0] multicycle_result;
|
|
|
|
if (RV32B != RV32BNone) begin : g_alu_rvb
|
|
|
|
/////////////////
|
|
// Bitcounting //
|
|
/////////////////
|
|
|
|
// The bit-counter structure computes the number of set bits in its operand. Partial results
|
|
// (from left to right) are needed to compute the control masks for computation of
|
|
// bcompress/bdecompress by the butterfly network, if implemented.
|
|
// For cpop, clz and ctz, only the end result is used.
|
|
|
|
logic zbe_op;
|
|
logic bitcnt_ctz;
|
|
logic bitcnt_clz;
|
|
logic bitcnt_cz;
|
|
logic [31:0] bitcnt_bits;
|
|
logic [31:0] bitcnt_mask_op;
|
|
logic [31:0] bitcnt_bit_mask;
|
|
logic [ 5:0] bitcnt_partial [32];
|
|
logic [31:0] bitcnt_partial_lsb_d;
|
|
logic [31:0] bitcnt_partial_msb_d;
|
|
|
|
|
|
assign bitcnt_ctz = operator_i == ALU_CTZ;
|
|
assign bitcnt_clz = operator_i == ALU_CLZ;
|
|
assign bitcnt_cz = bitcnt_ctz | bitcnt_clz;
|
|
assign bitcnt_result = bitcnt_partial[31];
|
|
|
|
// Bit-mask generation for clz and ctz:
|
|
// The bit mask is generated by spreading the lowest-order set bit in the operand to all
|
|
// higher order bits. The resulting mask is inverted to cover the lowest order zeros. In order
|
|
// to create the bit mask for leading zeros, the input operand needs to be reversed.
|
|
assign bitcnt_mask_op = bitcnt_clz ? operand_a_rev : operand_a_i;
|
|
|
|
always_comb begin
|
|
bitcnt_bit_mask = bitcnt_mask_op;
|
|
bitcnt_bit_mask |= bitcnt_bit_mask << 1;
|
|
bitcnt_bit_mask |= bitcnt_bit_mask << 2;
|
|
bitcnt_bit_mask |= bitcnt_bit_mask << 4;
|
|
bitcnt_bit_mask |= bitcnt_bit_mask << 8;
|
|
bitcnt_bit_mask |= bitcnt_bit_mask << 16;
|
|
bitcnt_bit_mask = ~bitcnt_bit_mask;
|
|
end
|
|
|
|
assign zbe_op = (operator_i == ALU_BCOMPRESS) | (operator_i == ALU_BDECOMPRESS);
|
|
|
|
always_comb begin
|
|
unique case (1'b1)
|
|
zbe_op: bitcnt_bits = operand_b_i;
|
|
bitcnt_cz: bitcnt_bits = bitcnt_bit_mask & ~bitcnt_mask_op; // clz / ctz
|
|
default: bitcnt_bits = operand_a_i; // cpop
|
|
endcase
|
|
end
|
|
|
|
// The parallel prefix counter is of the structure of a Brent-Kung Adder. In the first
|
|
// log2(width) stages, the sum of the n preceding bit lines is computed for the bit lines at
|
|
// positions 2**n-1 (power-of-two positions) where n denotes the current stage.
|
|
// In stage n=log2(width), the count for position width-1 (the MSB) is finished.
|
|
// For the intermediate values, an inverse adder tree then computes the bit counts for the bit
|
|
// lines at positions
|
|
// m = 2**(n-1) + i*2**(n-2), where i = [1 ... width / 2**(n-1)-1] and n = [log2(width) ... 2].
|
|
// Thus, at every subsequent stage the result of two previously unconnected sub-trees is
|
|
// summed, starting at the node summing bits [width/2-1 : 0] and [3*width/4-1: width/2]
|
|
// and moving to iteratively sum up all the sub-trees.
|
|
// The inverse adder tree thus features log2(width) - 1 stages the first of these stages is a
|
|
// single addition at position 3*width/4 - 1. It does not interfere with the last
|
|
// stage of the primary adder tree. These stages can thus be folded together, resulting in a
|
|
// total of 2*log2(width)-2 stages.
|
|
// For more details refer to R. Brent, H. T. Kung, "A Regular Layout for Parallel Adders",
|
|
// (1982).
|
|
// For a bitline at position p, only bits
|
|
// bitcnt_partial[max(i, such that p % log2(i) == 0)-1 : 0] are needed for generation of the
|
|
// butterfly network control signals. The adders in the intermediate value adder tree thus need
|
|
// not be full 5-bit adders. We leave the optimization to the synthesis tools.
|
|
//
|
|
// Consider the following 8-bit example for illustraton.
|
|
//
|
|
// let bitcnt_bits = 8'babcdefgh.
|
|
//
|
|
// a b c d e f g h
|
|
// | /: | /: | /: | /:
|
|
// |/ : |/ : |/ : |/ :
|
|
// stage 1: + : + : + : + :
|
|
// | : /: : | : /: :
|
|
// |,--+ : : |,--+ : :
|
|
// stage 2: + : : : + : : :
|
|
// | : | : /: : : :
|
|
// |,-----,--+ : : : : ^-primary adder tree
|
|
// stage 3: + : + : : : : : -------------------------
|
|
// : | /| /| /| /| /| : ,-intermediate adder tree
|
|
// : |/ |/ |/ |/ |/ : :
|
|
// stage 4 : + + + + + : :
|
|
// : : : : : : : :
|
|
// bitcnt_partial[i] 7 6 5 4 3 2 1 0
|
|
|
|
always_comb begin
|
|
bitcnt_partial = '{default: '0};
|
|
// stage 1
|
|
for (int unsigned i = 1; i < 32; i += 2) begin
|
|
bitcnt_partial[i] = {5'h0, bitcnt_bits[i]} + {5'h0, bitcnt_bits[i-1]};
|
|
end
|
|
// stage 2
|
|
for (int unsigned i = 3; i < 32; i += 4) begin
|
|
bitcnt_partial[i] = bitcnt_partial[i-2] + bitcnt_partial[i];
|
|
end
|
|
// stage 3
|
|
for (int unsigned i = 7; i < 32; i += 8) begin
|
|
bitcnt_partial[i] = bitcnt_partial[i-4] + bitcnt_partial[i];
|
|
end
|
|
// stage 4
|
|
for (int unsigned i = 15; i < 32; i += 16) begin
|
|
bitcnt_partial[i] = bitcnt_partial[i-8] + bitcnt_partial[i];
|
|
end
|
|
// stage 5
|
|
bitcnt_partial[31] = bitcnt_partial[15] + bitcnt_partial[31];
|
|
// ^- primary adder tree
|
|
// -------------------------------
|
|
// ,-intermediate value adder tree
|
|
bitcnt_partial[23] = bitcnt_partial[15] + bitcnt_partial[23];
|
|
|
|
// stage 6
|
|
for (int unsigned i = 11; i < 32; i += 8) begin
|
|
bitcnt_partial[i] = bitcnt_partial[i-4] + bitcnt_partial[i];
|
|
end
|
|
|
|
// stage 7
|
|
for (int unsigned i = 5; i < 32; i += 4) begin
|
|
bitcnt_partial[i] = bitcnt_partial[i-2] + bitcnt_partial[i];
|
|
end
|
|
// stage 8
|
|
bitcnt_partial[0] = {5'h0, bitcnt_bits[0]};
|
|
for (int unsigned i = 2; i < 32; i += 2) begin
|
|
bitcnt_partial[i] = bitcnt_partial[i-1] + {5'h0, bitcnt_bits[i]};
|
|
end
|
|
end
|
|
|
|
///////////////
|
|
// Min / Max //
|
|
///////////////
|
|
|
|
assign minmax_result = cmp_result ? operand_a_i : operand_b_i;
|
|
|
|
//////////
|
|
// Pack //
|
|
//////////
|
|
|
|
logic packu;
|
|
logic packh;
|
|
assign packu = operator_i == ALU_PACKU;
|
|
assign packh = operator_i == ALU_PACKH;
|
|
|
|
always_comb begin
|
|
unique case (1'b1)
|
|
packu: pack_result = {operand_b_i[31:16], operand_a_i[31:16]};
|
|
packh: pack_result = {16'h0, operand_b_i[7:0], operand_a_i[7:0]};
|
|
default: pack_result = {operand_b_i[15:0], operand_a_i[15:0]};
|
|
endcase
|
|
end
|
|
|
|
//////////
|
|
// Sext //
|
|
//////////
|
|
|
|
assign sext_result = (operator_i == ALU_SEXTB) ?
|
|
{ {24{operand_a_i[7]}}, operand_a_i[7:0]} : { {16{operand_a_i[15]}}, operand_a_i[15:0]};
|
|
|
|
/////////////////////////////
|
|
// Single-bit Instructions //
|
|
/////////////////////////////
|
|
|
|
always_comb begin
|
|
unique case (operator_i)
|
|
ALU_BSET: singlebit_result = operand_a_i | shift_result;
|
|
ALU_BCLR: singlebit_result = operand_a_i & ~shift_result;
|
|
ALU_BINV: singlebit_result = operand_a_i ^ shift_result;
|
|
default: singlebit_result = {31'h0, shift_result[0]}; // ALU_BEXT
|
|
endcase
|
|
end
|
|
|
|
////////////////////////////////////
|
|
// General Reverse and Or-combine //
|
|
////////////////////////////////////
|
|
|
|
// Only a subset of the general reverse and or-combine instructions are implemented in the
|
|
// balanced version of the B extension. Currently rev8 (shift_amt = 5'b11000) and orc.b
|
|
// (shift_amt = 5'b00111) are supported in the base extension.
|
|
|
|
logic [4:0] zbp_shift_amt;
|
|
logic gorc_op;
|
|
|
|
assign gorc_op = (operator_i == ALU_GORC);
|
|
assign zbp_shift_amt[2:0] =
|
|
(RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) ? shift_amt[2:0] : {3{shift_amt[0]}};
|
|
assign zbp_shift_amt[4:3] =
|
|
(RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) ? shift_amt[4:3] : {2{shift_amt[3]}};
|
|
|
|
always_comb begin
|
|
rev_result = operand_a_i;
|
|
|
|
if (zbp_shift_amt[0]) begin
|
|
rev_result = (gorc_op ? rev_result : 32'h0) |
|
|
((rev_result & 32'h5555_5555) << 1) |
|
|
((rev_result & 32'haaaa_aaaa) >> 1);
|
|
end
|
|
|
|
if (zbp_shift_amt[1]) begin
|
|
rev_result = (gorc_op ? rev_result : 32'h0) |
|
|
((rev_result & 32'h3333_3333) << 2) |
|
|
((rev_result & 32'hcccc_cccc) >> 2);
|
|
end
|
|
|
|
if (zbp_shift_amt[2]) begin
|
|
rev_result = (gorc_op ? rev_result : 32'h0) |
|
|
((rev_result & 32'h0f0f_0f0f) << 4) |
|
|
((rev_result & 32'hf0f0_f0f0) >> 4);
|
|
end
|
|
|
|
if (zbp_shift_amt[3]) begin
|
|
rev_result = ((RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) &&
|
|
gorc_op ? rev_result : 32'h0) |
|
|
((rev_result & 32'h00ff_00ff) << 8) |
|
|
((rev_result & 32'hff00_ff00) >> 8);
|
|
end
|
|
|
|
if (zbp_shift_amt[4]) begin
|
|
rev_result = ((RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) &&
|
|
gorc_op ? rev_result : 32'h0) |
|
|
((rev_result & 32'h0000_ffff) << 16) |
|
|
((rev_result & 32'hffff_0000) >> 16);
|
|
end
|
|
end
|
|
|
|
logic crc_hmode;
|
|
logic crc_bmode;
|
|
logic [31:0] clmul_result_rev;
|
|
|
|
if (RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) begin : gen_alu_rvb_otearlgrey_full
|
|
|
|
/////////////////////////
|
|
// Shuffle / Unshuffle //
|
|
/////////////////////////
|
|
|
|
localparam logic [31:0] SHUFFLE_MASK_L [4] =
|
|
'{32'h00ff_0000, 32'h0f00_0f00, 32'h3030_3030, 32'h4444_4444};
|
|
localparam logic [31:0] SHUFFLE_MASK_R [4] =
|
|
'{32'h0000_ff00, 32'h00f0_00f0, 32'h0c0c_0c0c, 32'h2222_2222};
|
|
|
|
localparam logic [31:0] FLIP_MASK_L [4] =
|
|
'{32'h2200_1100, 32'h0044_0000, 32'h4411_0000, 32'h1100_0000};
|
|
localparam logic [31:0] FLIP_MASK_R [4] =
|
|
'{32'h0088_0044, 32'h0000_2200, 32'h0000_8822, 32'h0000_0088};
|
|
|
|
logic [31:0] SHUFFLE_MASK_NOT [4];
|
|
for(genvar i = 0; i < 4; i++) begin : gen_shuffle_mask_not
|
|
assign SHUFFLE_MASK_NOT[i] = ~(SHUFFLE_MASK_L[i] | SHUFFLE_MASK_R[i]);
|
|
end
|
|
|
|
logic shuffle_flip;
|
|
assign shuffle_flip = operator_i == ALU_UNSHFL;
|
|
|
|
logic [3:0] shuffle_mode;
|
|
|
|
always_comb begin
|
|
shuffle_result = operand_a_i;
|
|
|
|
if (shuffle_flip) begin
|
|
shuffle_mode[3] = shift_amt[0];
|
|
shuffle_mode[2] = shift_amt[1];
|
|
shuffle_mode[1] = shift_amt[2];
|
|
shuffle_mode[0] = shift_amt[3];
|
|
end else begin
|
|
shuffle_mode = shift_amt[3:0];
|
|
end
|
|
|
|
if (shuffle_flip) begin
|
|
shuffle_result = (shuffle_result & 32'h8822_4411) |
|
|
((shuffle_result << 6) & FLIP_MASK_L[0]) |
|
|
((shuffle_result >> 6) & FLIP_MASK_R[0]) |
|
|
((shuffle_result << 9) & FLIP_MASK_L[1]) |
|
|
((shuffle_result >> 9) & FLIP_MASK_R[1]) |
|
|
((shuffle_result << 15) & FLIP_MASK_L[2]) |
|
|
((shuffle_result >> 15) & FLIP_MASK_R[2]) |
|
|
((shuffle_result << 21) & FLIP_MASK_L[3]) |
|
|
((shuffle_result >> 21) & FLIP_MASK_R[3]);
|
|
end
|
|
|
|
if (shuffle_mode[3]) begin
|
|
shuffle_result = (shuffle_result & SHUFFLE_MASK_NOT[0]) |
|
|
(((shuffle_result << 8) & SHUFFLE_MASK_L[0]) |
|
|
((shuffle_result >> 8) & SHUFFLE_MASK_R[0]));
|
|
end
|
|
if (shuffle_mode[2]) begin
|
|
shuffle_result = (shuffle_result & SHUFFLE_MASK_NOT[1]) |
|
|
(((shuffle_result << 4) & SHUFFLE_MASK_L[1]) |
|
|
((shuffle_result >> 4) & SHUFFLE_MASK_R[1]));
|
|
end
|
|
if (shuffle_mode[1]) begin
|
|
shuffle_result = (shuffle_result & SHUFFLE_MASK_NOT[2]) |
|
|
(((shuffle_result << 2) & SHUFFLE_MASK_L[2]) |
|
|
((shuffle_result >> 2) & SHUFFLE_MASK_R[2]));
|
|
end
|
|
if (shuffle_mode[0]) begin
|
|
shuffle_result = (shuffle_result & SHUFFLE_MASK_NOT[3]) |
|
|
(((shuffle_result << 1) & SHUFFLE_MASK_L[3]) |
|
|
((shuffle_result >> 1) & SHUFFLE_MASK_R[3]));
|
|
end
|
|
|
|
if (shuffle_flip) begin
|
|
shuffle_result = (shuffle_result & 32'h8822_4411) |
|
|
((shuffle_result << 6) & FLIP_MASK_L[0]) |
|
|
((shuffle_result >> 6) & FLIP_MASK_R[0]) |
|
|
((shuffle_result << 9) & FLIP_MASK_L[1]) |
|
|
((shuffle_result >> 9) & FLIP_MASK_R[1]) |
|
|
((shuffle_result << 15) & FLIP_MASK_L[2]) |
|
|
((shuffle_result >> 15) & FLIP_MASK_R[2]) |
|
|
((shuffle_result << 21) & FLIP_MASK_L[3]) |
|
|
((shuffle_result >> 21) & FLIP_MASK_R[3]);
|
|
end
|
|
end
|
|
|
|
//////////////
|
|
// Crossbar //
|
|
//////////////
|
|
// The crossbar permutation instructions xperm.[nbh] (Zbp) can be implemented using 8
|
|
// parallel 4-bit-wide, 8-input crossbars. Basically, we permute the 8 nibbles of operand_a_i
|
|
// based on operand_b_i.
|
|
|
|
// Generate selector indices and valid signals.
|
|
// - sel_n[x] indicates which nibble of operand_a_i is selected for output nibble x.
|
|
// - vld_n[x] indicates if the selection is valid.
|
|
logic [7:0][2:0] sel_n; // nibbles
|
|
logic [7:0] vld_n; // nibbles
|
|
logic [3:0][1:0] sel_b; // bytes
|
|
logic [3:0] vld_b; // bytes
|
|
logic [1:0][0:0] sel_h; // half words
|
|
logic [1:0] vld_h; // half words
|
|
|
|
// Per nibble, 3 bits are needed for the selection. Other bits must be zero.
|
|
// sel_n bit mask: 32'b0111_0111_0111_0111_0111_0111_0111_0111
|
|
// vld_n bit mask: 32'b1000_1000_1000_1000_1000_1000_1000_1000
|
|
for (genvar i = 0; i < 8; i++) begin : gen_sel_vld_n
|
|
assign sel_n[i] = operand_b_i[i*4 +: 3];
|
|
assign vld_n[i] = ~|operand_b_i[i*4 + 3 +: 1];
|
|
end
|
|
|
|
// Per byte, 2 bits are needed for the selection. Other bits must be zero.
|
|
// sel_b bit mask: 32'b0000_0011_0000_0011_0000_0011_0000_0011
|
|
// vld_b bit mask: 32'b1111_1100_1111_1100_1111_1100_1111_1100
|
|
for (genvar i = 0; i < 4; i++) begin : gen_sel_vld_b
|
|
assign sel_b[i] = operand_b_i[i*8 +: 2];
|
|
assign vld_b[i] = ~|operand_b_i[i*8 + 2 +: 6];
|
|
end
|
|
|
|
// Per half word, 1 bit is needed for the selection only. All other bits must be zero.
|
|
// sel_h bit mask: 32'b0000_0000_0000_0001_0000_0000_0000_0001
|
|
// vld_h bit mask: 32'b1111_1111_1111_1110_1111_1111_1111_1110
|
|
for (genvar i = 0; i < 2; i++) begin : gen_sel_vld_h
|
|
assign sel_h[i] = operand_b_i[i*16 +: 1];
|
|
assign vld_h[i] = ~|operand_b_i[i*16 + 1 +: 15];
|
|
end
|
|
|
|
// Convert selector indices and valid signals to control the nibble-based
|
|
// crossbar logic.
|
|
logic [7:0][2:0] sel;
|
|
logic [7:0] vld;
|
|
always_comb begin
|
|
unique case (operator_i)
|
|
ALU_XPERM_N: begin
|
|
// No conversion needed.
|
|
sel = sel_n;
|
|
vld = vld_n;
|
|
end
|
|
|
|
ALU_XPERM_B: begin
|
|
// Convert byte to nibble indicies.
|
|
for (int b = 0; b < 4; b++) begin
|
|
sel[b*2 + 0] = {sel_b[b], 1'b0};
|
|
sel[b*2 + 1] = {sel_b[b], 1'b1};
|
|
vld[b*2 +: 2] = {2{vld_b[b]}};
|
|
end
|
|
end
|
|
|
|
ALU_XPERM_H: begin
|
|
// Convert half-word to nibble indices.
|
|
for (int h = 0; h < 2; h++) begin
|
|
sel[h*4 + 0] = {sel_h[h], 2'b00};
|
|
sel[h*4 + 1] = {sel_h[h], 2'b01};
|
|
sel[h*4 + 2] = {sel_h[h], 2'b10};
|
|
sel[h*4 + 3] = {sel_h[h], 2'b11};
|
|
vld[h*4 +: 4] = {4{vld_h[h]}};
|
|
end
|
|
end
|
|
|
|
default: begin
|
|
// Tie valid to zero to disable the crossbar unless we need it.
|
|
sel = sel_n;
|
|
vld = '0;
|
|
end
|
|
endcase
|
|
end
|
|
|
|
// The actual nibble-based crossbar logic.
|
|
logic [7:0][3:0] val_n;
|
|
logic [7:0][3:0] xperm_n;
|
|
assign val_n = operand_a_i;
|
|
for (genvar i = 0; i < 8; i++) begin : gen_xperm_n
|
|
assign xperm_n[i] = vld[i] ? val_n[sel[i]] : '0;
|
|
end
|
|
assign xperm_result = xperm_n;
|
|
|
|
///////////////////////////////////////////////////
|
|
// Carry-less Multiply + Cyclic Redundancy Check //
|
|
///////////////////////////////////////////////////
|
|
|
|
// Carry-less multiplication can be understood as multiplication based on
|
|
// the addition interpreted as the bit-wise xor operation.
|
|
//
|
|
// Example: 1101 X 1011 = 1111111:
|
|
//
|
|
// 1011 X 1101
|
|
// -----------
|
|
// 1101
|
|
// xor 1101
|
|
// ---------
|
|
// 10111
|
|
// xor 0000
|
|
// ----------
|
|
// 010111
|
|
// xor 1101
|
|
// -----------
|
|
// 1111111
|
|
//
|
|
// Architectural details:
|
|
// A 32 x 32-bit array
|
|
// [ operand_b[i] ? (operand_a << i) : '0 for i in 0 ... 31 ]
|
|
// is generated. The entries of the array are pairwise 'xor-ed'
|
|
// together in a 5-stage binary tree.
|
|
//
|
|
//
|
|
// Cyclic Redundancy Check:
|
|
//
|
|
// CRC-32 (CRC-32/ISO-HDLC) and CRC-32C (CRC-32/ISCSI) are directly implemented. For
|
|
// documentation of the crc configuration (crc-polynomials, initialization, reflection, etc.)
|
|
// see http://reveng.sourceforge.net/crc-catalogue/all.htm
|
|
// A useful guide to crc arithmetic and algorithms is given here:
|
|
// http://www.piclist.com/techref/method/math/crcguide.html.
|
|
//
|
|
// The CRC operation solves the following equation using binary polynomial arithmetic:
|
|
//
|
|
// rev(rd)(x) = rev(rs1)(x) * x**n mod {1, P}(x)
|
|
//
|
|
// where P denotes lower 32 bits of the corresponding CRC polynomial, rev(a) the bit reversal
|
|
// of a, n = 8,16, or 32 for .b, .h, .w -variants. {a, b} denotes bit concatenation.
|
|
//
|
|
// Using barret reduction, one can show that
|
|
//
|
|
// M(x) mod P(x) = R(x) =
|
|
// (M(x) * x**n) & {deg(P(x)'{1'b1}}) ^ (M(x) x**-(deg(P(x) - n)) cx mu(x) cx P(x),
|
|
//
|
|
// Where mu(x) = polydiv(x**64, {1,P}) & 0xffffffff. Here, 'cx' refers to carry-less
|
|
// multiplication. Substituting rev(rd)(x) for R(x) and rev(rs1)(x) for M(x) and solving for
|
|
// rd(x) with P(x) a crc32 polynomial (deg(P(x)) = 32), we get
|
|
//
|
|
// rd = rev( (rev(rs1) << n) ^ ((rev(rs1) >> (32-n)) cx mu cx P)
|
|
// = (rs1 >> n) ^ rev(rev( (rs1 << (32-n)) cx rev(mu)) cx P)
|
|
// ^-- cycle 0--------------------^
|
|
// ^- cycle 1 -------------------------------------------^
|
|
//
|
|
// In the last step we used the fact that carry-less multiplication is bit-order agnostic:
|
|
// rev(a cx b) = rev(a) cx rev(b).
|
|
|
|
logic clmul_rmode;
|
|
logic clmul_hmode;
|
|
logic [31:0] clmul_op_a;
|
|
logic [31:0] clmul_op_b;
|
|
logic [31:0] operand_b_rev;
|
|
logic [31:0] clmul_and_stage[32];
|
|
logic [31:0] clmul_xor_stage1[16];
|
|
logic [31:0] clmul_xor_stage2[8];
|
|
logic [31:0] clmul_xor_stage3[4];
|
|
logic [31:0] clmul_xor_stage4[2];
|
|
|
|
logic [31:0] clmul_result_raw;
|
|
|
|
for (genvar i = 0; i < 32; i++) begin : gen_rev_operand_b
|
|
assign operand_b_rev[i] = operand_b_i[31-i];
|
|
end
|
|
|
|
assign clmul_rmode = operator_i == ALU_CLMULR;
|
|
assign clmul_hmode = operator_i == ALU_CLMULH;
|
|
|
|
// CRC
|
|
localparam logic [31:0] CRC32_POLYNOMIAL = 32'h04c1_1db7;
|
|
localparam logic [31:0] CRC32_MU_REV = 32'hf701_1641;
|
|
|
|
localparam logic [31:0] CRC32C_POLYNOMIAL = 32'h1edc_6f41;
|
|
localparam logic [31:0] CRC32C_MU_REV = 32'hdea7_13f1;
|
|
|
|
logic crc_op;
|
|
|
|
logic crc_cpoly;
|
|
|
|
logic [31:0] crc_operand;
|
|
logic [31:0] crc_poly;
|
|
logic [31:0] crc_mu_rev;
|
|
|
|
assign crc_op = (operator_i == ALU_CRC32C_W) | (operator_i == ALU_CRC32_W) |
|
|
(operator_i == ALU_CRC32C_H) | (operator_i == ALU_CRC32_H) |
|
|
(operator_i == ALU_CRC32C_B) | (operator_i == ALU_CRC32_B);
|
|
|
|
assign crc_cpoly = (operator_i == ALU_CRC32C_W) |
|
|
(operator_i == ALU_CRC32C_H) |
|
|
(operator_i == ALU_CRC32C_B);
|
|
|
|
assign crc_hmode = (operator_i == ALU_CRC32_H) | (operator_i == ALU_CRC32C_H);
|
|
assign crc_bmode = (operator_i == ALU_CRC32_B) | (operator_i == ALU_CRC32C_B);
|
|
|
|
assign crc_poly = crc_cpoly ? CRC32C_POLYNOMIAL : CRC32_POLYNOMIAL;
|
|
assign crc_mu_rev = crc_cpoly ? CRC32C_MU_REV : CRC32_MU_REV;
|
|
|
|
always_comb begin
|
|
unique case (1'b1)
|
|
crc_bmode: crc_operand = {operand_a_i[7:0], 24'h0};
|
|
crc_hmode: crc_operand = {operand_a_i[15:0], 16'h0};
|
|
default: crc_operand = operand_a_i;
|
|
endcase
|
|
end
|
|
|
|
// Select clmul input
|
|
always_comb begin
|
|
if (crc_op) begin
|
|
clmul_op_a = instr_first_cycle_i ? crc_operand : imd_val_q_i[0];
|
|
clmul_op_b = instr_first_cycle_i ? crc_mu_rev : crc_poly;
|
|
end else begin
|
|
clmul_op_a = clmul_rmode | clmul_hmode ? operand_a_rev : operand_a_i;
|
|
clmul_op_b = clmul_rmode | clmul_hmode ? operand_b_rev : operand_b_i;
|
|
end
|
|
end
|
|
|
|
for (genvar i = 0; i < 32; i++) begin : gen_clmul_and_op
|
|
assign clmul_and_stage[i] = clmul_op_b[i] ? clmul_op_a << i : '0;
|
|
end
|
|
|
|
for (genvar i = 0; i < 16; i++) begin : gen_clmul_xor_op_l1
|
|
assign clmul_xor_stage1[i] = clmul_and_stage[2*i] ^ clmul_and_stage[2*i+1];
|
|
end
|
|
|
|
for (genvar i = 0; i < 8; i++) begin : gen_clmul_xor_op_l2
|
|
assign clmul_xor_stage2[i] = clmul_xor_stage1[2*i] ^ clmul_xor_stage1[2*i+1];
|
|
end
|
|
|
|
for (genvar i = 0; i < 4; i++) begin : gen_clmul_xor_op_l3
|
|
assign clmul_xor_stage3[i] = clmul_xor_stage2[2*i] ^ clmul_xor_stage2[2*i+1];
|
|
end
|
|
|
|
for (genvar i = 0; i < 2; i++) begin : gen_clmul_xor_op_l4
|
|
assign clmul_xor_stage4[i] = clmul_xor_stage3[2*i] ^ clmul_xor_stage3[2*i+1];
|
|
end
|
|
|
|
assign clmul_result_raw = clmul_xor_stage4[0] ^ clmul_xor_stage4[1];
|
|
|
|
for (genvar i = 0; i < 32; i++) begin : gen_rev_clmul_result
|
|
assign clmul_result_rev[i] = clmul_result_raw[31-i];
|
|
end
|
|
|
|
// clmulr_result = rev(clmul(rev(a), rev(b)))
|
|
// clmulh_result = clmulr_result >> 1
|
|
always_comb begin
|
|
unique case (1'b1)
|
|
clmul_rmode: clmul_result = clmul_result_rev;
|
|
clmul_hmode: clmul_result = {1'b0, clmul_result_rev[31:1]};
|
|
default: clmul_result = clmul_result_raw;
|
|
endcase
|
|
end
|
|
end else begin : gen_alu_rvb_not_otearlgrey_full
|
|
assign shuffle_result = '0;
|
|
assign xperm_result = '0;
|
|
assign clmul_result = '0;
|
|
// support signals
|
|
assign clmul_result_rev = '0;
|
|
assign crc_bmode = '0;
|
|
assign crc_hmode = '0;
|
|
end
|
|
|
|
if (RV32B == RV32BFull) begin : gen_alu_rvb_full
|
|
|
|
///////////////
|
|
// Butterfly //
|
|
///////////////
|
|
|
|
// The butterfly / inverse butterfly network executing bcompress/bdecompress (zbe)
|
|
// instructions. For bdecompress, the control bits mask of a local left region is generated
|
|
// by the inverse of a n-bit left rotate and complement upon wrap (LROTC) operation by the
|
|
// number of ones in the deposit bitmask to the right of the segment. n hereby denotes the
|
|
// width of the according segment. The bitmask for a pertaining local right region is equal
|
|
// to the corresponding local left region. Bcompress uses an analogue inverse process.
|
|
// Consider the following 8-bit example. For details, see Hilewitz et al. "Fast Bit Gather,
|
|
// Bit Scatter and Bit Permuation Instructions for Commodity Microprocessors", (2008).
|
|
//
|
|
// The bcompress/bdecompress instructions are completed in 2 cycles. In the first cycle, the
|
|
// control bitmask is prepared by executing the parallel prefix bit count. In the second
|
|
// cycle, the bit swapping is executed according to the control masks.
|
|
|
|
// 8-bit example: (Hilewitz et al.)
|
|
// Consider the instruction bdecompress operand_a_i deposit_mask
|
|
// Let operand_a_i = 8'babcd_efgh
|
|
// deposit_mask = 8'b1010_1101
|
|
//
|
|
// control bitmask for stage 1:
|
|
// - number of ones in the right half of the deposit bitmask: 3
|
|
// - width of the segment: 4
|
|
// - control bitmask = ~LROTC(4'b0, 3)[3:0] = 4'b1000
|
|
//
|
|
// control bitmask: c3 c2 c1 c0 c3 c2 c1 c0
|
|
// 1 0 0 0 1 0 0 0
|
|
// <- L -----> <- R ----->
|
|
// operand_a_i a b c d e f g h
|
|
// :\ | | | /: | | |
|
|
// : +|---|--|-+ : | | |
|
|
// :/ | | | \: | | |
|
|
// stage 1 e b c d a f g h
|
|
// <L-> <R-> <L-> <R->
|
|
// control bitmask: c3 c2 c3 c2 c1 c0 c1 c0
|
|
// 1 1 1 1 1 0 1 0
|
|
// :\ :\ /: /: :\ | /: |
|
|
// : +:-+-:+ : : +|-+ : |
|
|
// :/ :/ \: \: :/ | \: |
|
|
// stage 2 c d e b g f a h
|
|
// L R L R L R L R
|
|
// control bitmask: c3 c3 c2 c2 c1 c1 c0 c0
|
|
// 1 1 0 0 1 1 0 0
|
|
// :\/: | | :\/: | |
|
|
// : : | | : : | |
|
|
// :/\: | | :/\: | |
|
|
// stage 3 d c e b f g a h
|
|
// & deposit bitmask: 1 0 1 0 1 1 0 1
|
|
// result: d 0 e 0 f g 0 h
|
|
|
|
logic [ 5:0] bitcnt_partial_q [32];
|
|
|
|
// first cycle
|
|
// Store partial bitcnts
|
|
for (genvar i = 0; i < 32; i++) begin : gen_bitcnt_reg_in_lsb
|
|
assign bitcnt_partial_lsb_d[i] = bitcnt_partial[i][0];
|
|
end
|
|
|
|
for (genvar i = 0; i < 16; i++) begin : gen_bitcnt_reg_in_b1
|
|
assign bitcnt_partial_msb_d[i] = bitcnt_partial[2*i+1][1];
|
|
end
|
|
|
|
for (genvar i = 0; i < 8; i++) begin : gen_bitcnt_reg_in_b2
|
|
assign bitcnt_partial_msb_d[16+i] = bitcnt_partial[4*i+3][2];
|
|
end
|
|
|
|
for (genvar i = 0; i < 4; i++) begin : gen_bitcnt_reg_in_b3
|
|
assign bitcnt_partial_msb_d[24+i] = bitcnt_partial[8*i+7][3];
|
|
end
|
|
|
|
for (genvar i = 0; i < 2; i++) begin : gen_bitcnt_reg_in_b4
|
|
assign bitcnt_partial_msb_d[28+i] = bitcnt_partial[16*i+15][4];
|
|
end
|
|
|
|
assign bitcnt_partial_msb_d[30] = bitcnt_partial[31][5];
|
|
assign bitcnt_partial_msb_d[31] = 1'b0; // unused
|
|
|
|
// Second cycle
|
|
// Load partial bitcnts
|
|
always_comb begin
|
|
bitcnt_partial_q = '{default: '0};
|
|
|
|
for (int unsigned i = 0; i < 32; i++) begin : gen_bitcnt_reg_out_lsb
|
|
bitcnt_partial_q[i][0] = imd_val_q_i[0][i];
|
|
end
|
|
|
|
for (int unsigned i = 0; i < 16; i++) begin : gen_bitcnt_reg_out_b1
|
|
bitcnt_partial_q[2*i+1][1] = imd_val_q_i[1][i];
|
|
end
|
|
|
|
for (int unsigned i = 0; i < 8; i++) begin : gen_bitcnt_reg_out_b2
|
|
bitcnt_partial_q[4*i+3][2] = imd_val_q_i[1][16+i];
|
|
end
|
|
|
|
for (int unsigned i = 0; i < 4; i++) begin : gen_bitcnt_reg_out_b3
|
|
bitcnt_partial_q[8*i+7][3] = imd_val_q_i[1][24+i];
|
|
end
|
|
|
|
for (int unsigned i = 0; i < 2; i++) begin : gen_bitcnt_reg_out_b4
|
|
bitcnt_partial_q[16*i+15][4] = imd_val_q_i[1][28+i];
|
|
end
|
|
|
|
bitcnt_partial_q[31][5] = imd_val_q_i[1][30];
|
|
end
|
|
|
|
logic [31:0] butterfly_mask_l[5];
|
|
logic [31:0] butterfly_mask_r[5];
|
|
logic [31:0] butterfly_mask_not[5];
|
|
logic [31:0] lrotc_stage [5]; // left rotate and complement upon wrap
|
|
|
|
// number of bits in local r = 32 / 2**(stage + 1) = 16/2**stage
|
|
`define _N(stg) (16 >> stg)
|
|
|
|
// bcompress / bdecompress control bit generation
|
|
for (genvar stg = 0; stg < 5; stg++) begin : gen_butterfly_ctrl_stage
|
|
// number of segs: 2** stg
|
|
for (genvar seg=0; seg<2**stg; seg++) begin : gen_butterfly_ctrl
|
|
|
|
assign lrotc_stage[stg][2*`_N(stg)*(seg+1)-1 : 2*`_N(stg)*seg] =
|
|
{{`_N(stg){1'b0}},{`_N(stg){1'b1}}} <<
|
|
bitcnt_partial_q[`_N(stg)*(2*seg+1)-1][$clog2(`_N(stg)):0];
|
|
|
|
assign butterfly_mask_l[stg][`_N(stg)*(2*seg+2)-1 : `_N(stg)*(2*seg+1)]
|
|
= ~lrotc_stage[stg][`_N(stg)*(2*seg+2)-1 : `_N(stg)*(2*seg+1)];
|
|
|
|
assign butterfly_mask_r[stg][`_N(stg)*(2*seg+1)-1 : `_N(stg)*(2*seg)]
|
|
= ~lrotc_stage[stg][`_N(stg)*(2*seg+2)-1 : `_N(stg)*(2*seg+1)];
|
|
|
|
assign butterfly_mask_l[stg][`_N(stg)*(2*seg+1)-1 : `_N(stg)*(2*seg)] = '0;
|
|
assign butterfly_mask_r[stg][`_N(stg)*(2*seg+2)-1 : `_N(stg)*(2*seg+1)] = '0;
|
|
end
|
|
end
|
|
`undef _N
|
|
|
|
for (genvar stg = 0; stg < 5; stg++) begin : gen_butterfly_not
|
|
assign butterfly_mask_not[stg] =
|
|
~(butterfly_mask_l[stg] | butterfly_mask_r[stg]);
|
|
end
|
|
|
|
always_comb begin
|
|
butterfly_result = operand_a_i;
|
|
|
|
butterfly_result = butterfly_result & butterfly_mask_not[0] |
|
|
((butterfly_result & butterfly_mask_l[0]) >> 16)|
|
|
((butterfly_result & butterfly_mask_r[0]) << 16);
|
|
|
|
butterfly_result = butterfly_result & butterfly_mask_not[1] |
|
|
((butterfly_result & butterfly_mask_l[1]) >> 8)|
|
|
((butterfly_result & butterfly_mask_r[1]) << 8);
|
|
|
|
butterfly_result = butterfly_result & butterfly_mask_not[2] |
|
|
((butterfly_result & butterfly_mask_l[2]) >> 4)|
|
|
((butterfly_result & butterfly_mask_r[2]) << 4);
|
|
|
|
butterfly_result = butterfly_result & butterfly_mask_not[3] |
|
|
((butterfly_result & butterfly_mask_l[3]) >> 2)|
|
|
((butterfly_result & butterfly_mask_r[3]) << 2);
|
|
|
|
butterfly_result = butterfly_result & butterfly_mask_not[4] |
|
|
((butterfly_result & butterfly_mask_l[4]) >> 1)|
|
|
((butterfly_result & butterfly_mask_r[4]) << 1);
|
|
|
|
butterfly_result = butterfly_result & operand_b_i;
|
|
end
|
|
|
|
always_comb begin
|
|
invbutterfly_result = operand_a_i & operand_b_i;
|
|
|
|
invbutterfly_result = invbutterfly_result & butterfly_mask_not[4] |
|
|
((invbutterfly_result & butterfly_mask_l[4]) >> 1)|
|
|
((invbutterfly_result & butterfly_mask_r[4]) << 1);
|
|
|
|
invbutterfly_result = invbutterfly_result & butterfly_mask_not[3] |
|
|
((invbutterfly_result & butterfly_mask_l[3]) >> 2)|
|
|
((invbutterfly_result & butterfly_mask_r[3]) << 2);
|
|
|
|
invbutterfly_result = invbutterfly_result & butterfly_mask_not[2] |
|
|
((invbutterfly_result & butterfly_mask_l[2]) >> 4)|
|
|
((invbutterfly_result & butterfly_mask_r[2]) << 4);
|
|
|
|
invbutterfly_result = invbutterfly_result & butterfly_mask_not[1] |
|
|
((invbutterfly_result & butterfly_mask_l[1]) >> 8)|
|
|
((invbutterfly_result & butterfly_mask_r[1]) << 8);
|
|
|
|
invbutterfly_result = invbutterfly_result & butterfly_mask_not[0] |
|
|
((invbutterfly_result & butterfly_mask_l[0]) >> 16)|
|
|
((invbutterfly_result & butterfly_mask_r[0]) << 16);
|
|
end
|
|
end else begin : gen_alu_rvb_not_full
|
|
logic [31:0] unused_imd_val_q_1;
|
|
assign unused_imd_val_q_1 = imd_val_q_i[1];
|
|
assign butterfly_result = '0;
|
|
assign invbutterfly_result = '0;
|
|
// support signals
|
|
assign bitcnt_partial_lsb_d = '0;
|
|
assign bitcnt_partial_msb_d = '0;
|
|
end
|
|
|
|
//////////////////////////////////////
|
|
// Multicycle Bitmanip Instructions //
|
|
//////////////////////////////////////
|
|
// Ternary instructions + Shift Rotations + Bit Compress/Decompress + CRC
|
|
// For ternary instructions (zbt), operand_a_i is tied to rs1 in the first cycle and rs3 in the
|
|
// second cycle. operand_b_i is always tied to rs2.
|
|
|
|
always_comb begin
|
|
unique case (operator_i)
|
|
ALU_CMOV: begin
|
|
multicycle_result = (operand_b_i == 32'h0) ? operand_a_i : imd_val_q_i[0];
|
|
imd_val_d_o = '{operand_a_i, 32'h0};
|
|
if (instr_first_cycle_i) begin
|
|
imd_val_we_o = 2'b01;
|
|
end else begin
|
|
imd_val_we_o = 2'b00;
|
|
end
|
|
end
|
|
|
|
ALU_CMIX: begin
|
|
multicycle_result = imd_val_q_i[0] | bwlogic_and_result;
|
|
imd_val_d_o = '{bwlogic_and_result, 32'h0};
|
|
if (instr_first_cycle_i) begin
|
|
imd_val_we_o = 2'b01;
|
|
end else begin
|
|
imd_val_we_o = 2'b00;
|
|
end
|
|
end
|
|
|
|
ALU_FSR, ALU_FSL,
|
|
ALU_ROL, ALU_ROR: begin
|
|
if (shift_amt[4:0] == 5'h0) begin
|
|
multicycle_result = shift_amt[5] ? operand_a_i : imd_val_q_i[0];
|
|
end else begin
|
|
multicycle_result = imd_val_q_i[0] | shift_result;
|
|
end
|
|
imd_val_d_o = '{shift_result, 32'h0};
|
|
if (instr_first_cycle_i) begin
|
|
imd_val_we_o = 2'b01;
|
|
end else begin
|
|
imd_val_we_o = 2'b00;
|
|
end
|
|
end
|
|
|
|
ALU_CRC32_W, ALU_CRC32C_W,
|
|
ALU_CRC32_H, ALU_CRC32C_H,
|
|
ALU_CRC32_B, ALU_CRC32C_B: begin
|
|
if (RV32B == RV32BOTEarlGrey || RV32B == RV32BFull) begin
|
|
unique case (1'b1)
|
|
crc_bmode: multicycle_result = clmul_result_rev ^ (operand_a_i >> 8);
|
|
crc_hmode: multicycle_result = clmul_result_rev ^ (operand_a_i >> 16);
|
|
default: multicycle_result = clmul_result_rev;
|
|
endcase
|
|
imd_val_d_o = '{clmul_result_rev, 32'h0};
|
|
if (instr_first_cycle_i) begin
|
|
imd_val_we_o = 2'b01;
|
|
end else begin
|
|
imd_val_we_o = 2'b00;
|
|
end
|
|
end else begin
|
|
imd_val_d_o = '{operand_a_i, 32'h0};
|
|
imd_val_we_o = 2'b00;
|
|
multicycle_result = '0;
|
|
end
|
|
end
|
|
|
|
ALU_BCOMPRESS, ALU_BDECOMPRESS: begin
|
|
if (RV32B == RV32BFull) begin
|
|
multicycle_result = (operator_i == ALU_BDECOMPRESS) ? butterfly_result :
|
|
invbutterfly_result;
|
|
imd_val_d_o = '{bitcnt_partial_lsb_d, bitcnt_partial_msb_d};
|
|
if (instr_first_cycle_i) begin
|
|
imd_val_we_o = 2'b11;
|
|
end else begin
|
|
imd_val_we_o = 2'b00;
|
|
end
|
|
end else begin
|
|
imd_val_d_o = '{operand_a_i, 32'h0};
|
|
imd_val_we_o = 2'b00;
|
|
multicycle_result = '0;
|
|
end
|
|
end
|
|
|
|
default: begin
|
|
imd_val_d_o = '{operand_a_i, 32'h0};
|
|
imd_val_we_o = 2'b00;
|
|
multicycle_result = '0;
|
|
end
|
|
endcase
|
|
end
|
|
|
|
|
|
end else begin : g_no_alu_rvb
|
|
logic [31:0] unused_imd_val_q[2];
|
|
assign unused_imd_val_q = imd_val_q_i;
|
|
logic [31:0] unused_butterfly_result;
|
|
assign unused_butterfly_result = butterfly_result;
|
|
logic [31:0] unused_invbutterfly_result;
|
|
assign unused_invbutterfly_result = invbutterfly_result;
|
|
// RV32B result signals
|
|
assign bitcnt_result = '0;
|
|
assign minmax_result = '0;
|
|
assign pack_result = '0;
|
|
assign sext_result = '0;
|
|
assign singlebit_result = '0;
|
|
assign rev_result = '0;
|
|
assign shuffle_result = '0;
|
|
assign xperm_result = '0;
|
|
assign butterfly_result = '0;
|
|
assign invbutterfly_result = '0;
|
|
assign clmul_result = '0;
|
|
assign multicycle_result = '0;
|
|
// RV32B support signals
|
|
assign imd_val_d_o = '{default: '0};
|
|
assign imd_val_we_o = '{default: '0};
|
|
end
|
|
|
|
////////////////
|
|
// Result mux //
|
|
////////////////
|
|
|
|
always_comb begin
|
|
result_o = '0;
|
|
|
|
unique case (operator_i)
|
|
// Bitwise Logic Operations (negate: RV32B)
|
|
ALU_XOR, ALU_XNOR,
|
|
ALU_OR, ALU_ORN,
|
|
ALU_AND, ALU_ANDN: result_o = bwlogic_result;
|
|
|
|
// Adder Operations
|
|
ALU_ADD, ALU_SUB,
|
|
// RV32B
|
|
ALU_SH1ADD, ALU_SH2ADD,
|
|
ALU_SH3ADD: result_o = adder_result;
|
|
|
|
// Shift Operations
|
|
ALU_SLL, ALU_SRL,
|
|
ALU_SRA,
|
|
// RV32B
|
|
ALU_SLO, ALU_SRO: result_o = shift_result;
|
|
|
|
// Shuffle Operations (RV32B)
|
|
ALU_SHFL, ALU_UNSHFL: result_o = shuffle_result;
|
|
|
|
// Crossbar Permutation Operations (RV32B)
|
|
ALU_XPERM_N, ALU_XPERM_B, ALU_XPERM_H: result_o = xperm_result;
|
|
|
|
// Comparison Operations
|
|
ALU_EQ, ALU_NE,
|
|
ALU_GE, ALU_GEU,
|
|
ALU_LT, ALU_LTU,
|
|
ALU_SLT, ALU_SLTU: result_o = {31'h0,cmp_result};
|
|
|
|
// MinMax Operations (RV32B)
|
|
ALU_MIN, ALU_MAX,
|
|
ALU_MINU, ALU_MAXU: result_o = minmax_result;
|
|
|
|
// Bitcount Operations (RV32B)
|
|
ALU_CLZ, ALU_CTZ,
|
|
ALU_CPOP: result_o = {26'h0, bitcnt_result};
|
|
|
|
// Pack Operations (RV32B)
|
|
ALU_PACK, ALU_PACKH,
|
|
ALU_PACKU: result_o = pack_result;
|
|
|
|
// Sign-Extend (RV32B)
|
|
ALU_SEXTB, ALU_SEXTH: result_o = sext_result;
|
|
|
|
// Ternary Bitmanip Operations (RV32B)
|
|
ALU_CMIX, ALU_CMOV,
|
|
ALU_FSL, ALU_FSR,
|
|
// Rotate Shift (RV32B)
|
|
ALU_ROL, ALU_ROR,
|
|
// Cyclic Redundancy Checks (RV32B)
|
|
ALU_CRC32_W, ALU_CRC32C_W,
|
|
ALU_CRC32_H, ALU_CRC32C_H,
|
|
ALU_CRC32_B, ALU_CRC32C_B,
|
|
// Bit Compress / Decompress (RV32B)
|
|
ALU_BCOMPRESS, ALU_BDECOMPRESS: result_o = multicycle_result;
|
|
|
|
// Single-Bit Bitmanip Operations (RV32B)
|
|
ALU_BSET, ALU_BCLR,
|
|
ALU_BINV, ALU_BEXT: result_o = singlebit_result;
|
|
|
|
// General Reverse / Or-combine (RV32B)
|
|
ALU_GREV, ALU_GORC: result_o = rev_result;
|
|
|
|
// Bit Field Place (RV32B)
|
|
ALU_BFP: result_o = bfp_result;
|
|
|
|
// Carry-less Multiply Operations (RV32B)
|
|
ALU_CLMUL, ALU_CLMULR,
|
|
ALU_CLMULH: result_o = clmul_result;
|
|
|
|
default: ;
|
|
endcase
|
|
end
|
|
|
|
logic unused_shift_amt_compl;
|
|
assign unused_shift_amt_compl = shift_amt_compl[5];
|
|
|
|
endmodule
|