colin
/
Hazard3
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

First pass at implementing bitmanip. Breaks CXXRTL. Ooop

This commit is contained in:
Luke Wren 2021-11-25 22:58:49 +00:00
parent ed6b6a3660
commit 58c20a39d0
8 changed files with 251 additions and 50 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

169
hdl/arith/hazard3_alu.v
View File

@ -18,18 +18,26 @@
`default_nettype none
module hazard3_alu #(
parameter W_DATA = 32
`include "hazard3_config.vh"
,
`include "hazard3_width_const.vh"
) (
input wire [3:0] aluop,
input wire [W_DATA-1:0] op_a,
input wire [W_DATA-1:0] op_b,
output reg [W_DATA-1:0] result,
output wire [W_DATA-1:0] result_add, // for load/stores
output wire [W_DATA-1:0] result_add,
output wire cmp
);
`include "hazard3_ops.vh"
// ----------------------------------------------------------------------------
// Fiddle around with add/sub, comparisons etc (all related).
// This adder is exposed directly on the result_add port, since it may be used
// for load/store addresses.
function msb;
input [W_DATA-1:0] x;
begin
@ -37,59 +45,168 @@ begin
end
endfunction
wire sub = aluop != ALUOP_ADD;
wire [W_DATA-1:0] sum = op_a + (op_b ^ {W_DATA{sub}}) + sub;
wire sub = !(aluop == ALUOP_ADD || (|EXTENSION_ZBA && (
aluop == ALUOP_ADD_SH1 || aluop == ALUOP_ADD_SH2 || aluop == ALUOP_ADD_SH3
)));
wire inv_op_b = sub || (|EXTENSION_ZBB && (
aluop == ALUOP_ANDN || aluop == ALUOP_ORN || aluop == ALUOP_XNOR
));
wire [W_DATA-1:0] op_a_shifted =
|EXTENSION_ZBA && aluop == ALUOP_ADD_SH1 ? op_a << 1 :
|EXTENSION_ZBA && aluop == ALUOP_ADD_SH2 ? op_a << 2 :
|EXTENSION_ZBA && aluop == ALUOP_ADD_SH3 ? op_a << 3 : op_a;
wire [W_DATA-1:0] op_b_inv = op_b ^ {W_DATA{inv_op_b}};
wire [W_DATA-1:0] sum = op_a_shifted + op_b_inv + sub;
wire [W_DATA-1:0] op_xor = op_a ^ op_b;
wire cmp_is_unsigned = aluop == ALUOP_LTU ||
|EXTENSION_ZBB && aluop == ALUOP_MAXU ||
|EXTENSION_ZBB && aluop == ALUOP_MINU;
wire lt = msb(op_a) == msb(op_b) ? msb(sum) :
aluop == ALUOP_LTU ? msb(op_b) :
cmp_is_unsigned ? msb(op_b) :
msb(op_a) ;
assign cmp = aluop == ALUOP_SUB ? |op_xor : lt;
assign result_add = sum;
// ----------------------------------------------------------------------------
// Separate units for shift, ctz etc
wire [W_DATA-1:0] shift_dout;
wire shift_right_nleft = aluop == ALUOP_SRL || aluop == ALUOP_SRA;
wire shift_right_nleft = aluop == ALUOP_SRL || aluop == ALUOP_SRA ||
|EXTENSION_ZBB && aluop == ALUOP_ROR ||
|EXTENSION_ZBS && aluop == ALUOP_BEXT;
wire shift_arith = aluop == ALUOP_SRA;
wire shift_rotate = |EXTENSION_ZBB & (aluop == ALUOP_ROR || aluop == ALUOP_ROL);
hazard3_shift_barrel #(
.W_DATA(W_DATA),
.W_SHAMT(5)
`include "hazard3_config_inst.vh"
) shifter (
.din(op_a),
.shamt(op_b[4:0]),
.right_nleft(shift_right_nleft),
.arith(shift_arith),
.dout(shift_dout)
.din (op_a),
.shamt (op_b[4:0]),
.right_nleft (shift_right_nleft),
.rotate (shift_rotate),
.arith (shift_arith),
.dout (shift_dout)
);
// We can implement all bitwise ops with 1 LUT4/bit total, since each result bit
// uses only two operand bits. Much better than feeding each into main mux tree.
reg [W_DATA-1:0] op_a_rev;
always @ (*) begin: rev_op_a
integer i;
for (i = 0; i < W_DATA; i = i + 1) begin
op_a_rev[i] = op_a[W_DATA - 1 - i];
end
end
// "leading" means starting at MSB. This is an LSB-first priority encoder, so
// ""leading" is reversed and "trailing" is not.
wire [W_DATA-1:0] ctz_search_mask = aluop == ALUOP_CLZ ? op_a_rev : op_a;
wire [W_SHAMT:0] ctz_clz;
hazard3_priority_encode #(
.W_REQ (W_DATA)
) ctz_priority_encode (
.req (ctz_search_mask),
.gnt (ctz_clz[W_SHAMT-1:0])
);
// Special case: all-zeroes returns XLEN
assign ctz_clz[W_SHAMT] = ~|op_a;
reg [W_SHAMT:0] cpop;
always @ (*) begin: cpop_count
integer i;
cpop = {W_SHAMT+1{1'b0}};
for (i = 0; i < W_DATA; i = i + 1) begin
cpop = cpop + op_a[i];
end
end
reg [2*W_DATA-1:0] clmul;
always @ (*) begin: clmul_mul
integer i;
clmul = {2*W_DATA{1'b0}};
for (i = 0; i < W_DATA; i = i + 1) begin
clmul = clmul ^ (({{W_DATA{1'b0}}, op_a} << i) & {2*W_DATA{op_b[i]}});
end
end
// ----------------------------------------------------------------------------
// Output mux, with simple operations inline
// iCE40: We can implement all bitwise ops with 1 LUT4/bit total, since each
// result bit uses only two operand bits. Much better than feeding each into
// main mux tree. Doesn't matter for big-LUT FPGAs or for implementations with
// bitmanip extensions enabled.
reg [W_DATA-1:0] bitwise;
always @ (*) begin: bitwise_ops
case (aluop[1:0])
ALUOP_AND[1:0]: bitwise = op_a & op_b;
ALUOP_OR[1:0]: bitwise = op_a | op_b;
default: bitwise = op_a ^ op_b;
ALUOP_AND[1:0]: bitwise = op_a & op_b_inv;
ALUOP_OR[1:0]: bitwise = op_a | op_b_inv;
default: bitwise = op_a ^ op_b_inv;
endcase
end
wire [W_DATA-1:0] zbs_mask = {{W_DATA-1{1'b0}}, 1'b1} << op_b[W_SHAMT-1:0];
always @ (*) begin
case (aluop)
ALUOP_ADD: begin result = sum; end
ALUOP_SUB: begin result = sum; end
ALUOP_LT: begin result = {{W_DATA-1{1'b0}}, lt}; end
ALUOP_LTU: begin result = {{W_DATA-1{1'b0}}, lt}; end
ALUOP_SRL: begin result = shift_dout; end
ALUOP_SRA: begin result = shift_dout; end
ALUOP_SLL: begin result = shift_dout; end
casez ({|EXTENSION_ZBA, |EXTENSION_ZBB, EXTENSION_ZBC, |EXTENSION_ZBS, aluop})
// Base ISA
{4'bzzzz, ALUOP_ADD }: result = sum;
{4'bzzzz, ALUOP_SUB }: result = sum;
{4'bzzzz, ALUOP_LT }: result = {{W_DATA-1{1'b0}}, lt};
{4'bzzzz, ALUOP_LTU }: result = {{W_DATA-1{1'b0}}, lt};
{4'bzzzz, ALUOP_SRL }: result = shift_dout;
{4'bzzzz, ALUOP_SRA }: result = shift_dout;
{4'bzzzz, ALUOP_SLL }: result = shift_dout;
{4'bzzzz, ALUOP_SLL }: result = shift_dout;
// Zba
{4'b1zzz, ALUOP_ADD_SH1}: result = sum;
{4'b1zzz, ALUOP_ADD_SH2}: result = sum;
{4'b1zzz, ALUOP_ADD_SH3}: result = sum;
// Zbb
{4'bz1zz, ALUOP_ANDN }: result = bitwise;
{4'bz1zz, ALUOP_ORN }: result = bitwise;
{4'bz1zz, ALUOP_XNOR }: result = bitwise;
{4'bz1zz, ALUOP_CLZ }: result = ctz_clz;
{4'bz1zz, ALUOP_CTZ }: result = ctz_clz;
{4'bz1zz, ALUOP_CPOP }: result = cpop;
{4'bz1zz, ALUOP_MAX }: result = lt ? op_b : op_a;
{4'bz1zz, ALUOP_MAXU }: result = lt ? op_b : op_a;
{4'bz1zz, ALUOP_MIN }: result = lt ? op_a : op_b;
{4'bz1zz, ALUOP_MINU }: result = lt ? op_a : op_b;
{4'bz1zz, ALUOP_SEXT_B }: result = {{W_DATA-8{op_a[7]}}, op_a[7:0]};
{4'bz1zz, ALUOP_SEXT_H }: result = {{W_DATA-16{op_a[15]}}, op_a[15:0]};
{4'bz1zz, ALUOP_ZEXT_H }: result = {{W_DATA-16{1'b0}}, op_a[15:0]};
{4'bz1zz, ALUOP_ORC_B }: result = {{8{|op[31:24]}}, {8{|op[23:16]}}, {8{|op[15:8]}}, {8{|op[7:0]}}};
{4'bz1zz, ALUOP_REV8 }: result = {op[7:0], op[15:8], op[23:16], op[31:24]};
{4'bz1zz, ALUOP_ROL }: result = shift_dout;
{4'bz1zz, ALUOP_ROR }: result = shift_dout;
// Zbc
{4'bzz1z, ALUOP_CLMUL }: result = clmul[W_DATA-1:0];
{4'bzz1z, ALUOP_CLMULH }: result = clmul[2*W_DATA-1:W_DATA];
{4'bzz1z, ALUOP_CLMULR }: result = clmul[2*W_DATA-2:W_DATA-1];
// Zbs
{4'bzzz1, ALUOP_BCLR }: result = op_a & ~zbs_mask;
{4'bzzz1, ALUOP_BSET }: result = op_a | zbs_mask;
{4'bzzz1, ALUOP_BINV }: result = op_a ^ ~zbs_mask;
{4'bzzz1, ALUOP_BEXT }: result = {{W_DATA-1{1'b0}}, shift_dout[0]};
default: begin result = bitwise; end
endcase
end
// ----------------------------------------------------------------------------
// Properties for base-ISA instructions
`ifdef FORMAL
`ifndef RISCV_FORMAL
// Really we're just interested in the shifts and comparisons, as these are

15
hdl/arith/hazard3_shift_barrel.v
View File

@ -22,23 +22,25 @@
`default_nettype none
module hazard3_shift_barrel #(
parameter W_DATA = 32,
parameter W_SHAMT = 5
`include "hazard3_config.vh"
,
`include "hazard3_width_const.vh"
) (
input wire [W_DATA-1:0] din,
input wire [W_SHAMT-1:0] shamt,
input wire right_nleft,
input wire rotate,
input wire arith,
output reg [W_DATA-1:0] dout
);
integer i;
reg [W_DATA-1:0] din_rev;
reg [W_DATA-1:0] shift_accum;
wire sext = arith && din_rev[0]; // haha
always @ (*) begin
always @ (*) begin: shift
integer i;
for (i = 0; i < W_DATA; i = i + 1)
din_rev[i] = right_nleft ? din[W_DATA - 1 - i] : din[i];
@ -46,7 +48,8 @@ always @ (*) begin
for (i = 0; i < W_SHAMT; i = i + 1) begin
if (shamt[i]) begin
shift_accum = (shift_accum << (1 << i)) |
({W_DATA{sext}} & ~({W_DATA{1'b1}} << (1 << i)));
({W_DATA{sext}} & ~({W_DATA{1'b1}} << (1 << i))) |
({W_DATA{rotate && |EXTENSION_ZBB}} & (shift_accum >> (W_DATA - (1 << i))));
end
end

12
hdl/hazard3_config.vh
View File

@ -37,6 +37,18 @@ parameter EXTENSION_C = 1,
// EXTENSION_M: Support for hardware multiply/divide/modulo instructions
parameter EXTENSION_M = 1,
// EXTENSION_ZBA: Support for Zba address generation instructions
parameter EXTENSION_ZBA = 1,
// EXTENSION_ZBB: Support for Zbb basic bit manipulation instructions
parameter EXTENSION_ZBB = 1,
// EXTENSION_ZBC: Support for Zbc carry-less multiplication instructions
parameter EXTENSION_ZBC = 1,
// EXTENSION_ZBS: Support for Zbs single-bit manipulation instructions
parameter EXTENSION_ZBS = 1,
// CSR_M_MANDATORY: Bare minimum CSR support e.g. misa. Spec says must = 1 if
// CSRs are present, but I won't tell anyone.
parameter CSR_M_MANDATORY = 1,

4
hdl/hazard3_config_inst.vh
View File

@ -6,6 +6,10 @@
.MTVEC_INIT (MTVEC_INIT),
.EXTENSION_C (EXTENSION_C),
.EXTENSION_M (EXTENSION_M),
.EXTENSION_ZBA (EXTENSION_ZBA),
.EXTENSION_ZBB (EXTENSION_ZBB),
.EXTENSION_ZBC (EXTENSION_ZBC),
.EXTENSION_ZBS (EXTENSION_ZBS),
.CSR_M_MANDATORY (CSR_M_MANDATORY),
.CSR_M_TRAP (CSR_M_TRAP),
.CSR_COUNTER (CSR_COUNTER),

4
hdl/hazard3_core.v
View File

@ -345,7 +345,9 @@ always @ (*) begin
x_op_b = x_rs2_bypass;
end
hazard3_alu alu (
hazard3_alu #(
`include "hazard3_config_inst.vh"
) alu (
.aluop (d_aluop),
.op_a (x_op_a),
.op_b (x_op_b),

32
hdl/hazard3_decode.v
View File

@ -242,6 +242,38 @@ always @ (*) begin
RV_DIVU: if (EXTENSION_M) begin d_aluop = ALUOP_MULDIV; d_mulop = M_OP_DIVU; end else begin d_invalid_32bit = 1'b1; end
RV_REM: if (EXTENSION_M) begin d_aluop = ALUOP_MULDIV; d_mulop = M_OP_REM; end else begin d_invalid_32bit = 1'b1; end
RV_REMU: if (EXTENSION_M) begin d_aluop = ALUOP_MULDIV; d_mulop = M_OP_REMU; end else begin d_invalid_32bit = 1'b1; end
RV_SH1ADD: if (EXTENSION_ZBA) begin d_aluop = ALUOP_SH1ADD; end else begin d_invalid_32bit = 1'b1; end
RV_SH2ADD: if (EXTENSION_ZBA) begin d_aluop = ALUOP_SH2ADD; end else begin d_invalid_32bit = 1'b1; end
RV_SH3ADD: if (EXTENSION_ZBA) begin d_aluop = ALUOP_SH3ADD; end else begin d_invalid_32bit = 1'b1; end
RV_ANDN: if (EXTENSION_ZBB) begin d_aluop = ALUOP_ANDN; end else begin d_invalid_32bit = 1'b1; end
RV_CLZ: if (EXTENSION_ZBB) begin d_aluop = ALUOP_CLZ; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_CPOP: if (EXTENSION_ZBB) begin d_aluop = ALUOP_CPOP; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_CTZ: if (EXTENSION_ZBB) begin d_aluop = ALUOP_CTZ; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_MAX: if (EXTENSION_ZBB) begin d_aluop = ALUOP_MAX; end else begin d_invalid_32bit = 1'b1; end
RV_MAXU: if (EXTENSION_ZBB) begin d_aluop = ALUOP_MAXU; end else begin d_invalid_32bit = 1'b1; end
RV_MIN: if (EXTENSION_ZBB) begin d_aluop = ALUOP_MIN; end else begin d_invalid_32bit = 1'b1; end
RV_MINU: if (EXTENSION_ZBB) begin d_aluop = ALUOP_MINU; end else begin d_invalid_32bit = 1'b1; end
RV_ORC_B: if (EXTENSION_ZBB) begin d_aluop = ALUOP_ORC_B; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_ORN: if (EXTENSION_ZBB) begin d_aluop = ALUOP_ORN; end else begin d_invalid_32bit = 1'b1; end
RV_REV8: if (EXTENSION_ZBB) begin d_aluop = ALUOP_REV8; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_ROL: if (EXTENSION_ZBB) begin d_aluop = ALUOP_ROL; end else begin d_invalid_32bit = 1'b1; end
RV_ROR: if (EXTENSION_ZBB) begin d_aluop = ALUOP_ROR; end else begin d_invalid_32bit = 1'b1; end
RV_RORI: if (EXTENSION_ZBB) begin d_aluop = ALUOP_ROR; d_rs2 = X0; d_imm = d_imm_i; d_alusrc_b = ALUSRCB_IMM; end else begin d_invalid_32bit = 1'b1; end
RV_SEXT_B: if (EXTENSION_ZBB) begin d_aluop = ALUOP_SEXT_B; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_SEXT_H: if (EXTENSION_ZBB) begin d_aluop = ALUOP_SEXT_H; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_XNOR: if (EXTENSION_ZBB) begin d_aluop = ALUOP_XNOR; end else begin d_invalid_32bit = 1'b1; end
RV_ZEXT_H: if (EXTENSION_ZBB) begin d_aluop = ALUOP_ZEXT_H; d_rs2 = X0; end else begin d_invalid_32bit = 1'b1; end
RV_CLMUL: if (EXTENSION_ZBC) begin d_aluop = ALUOP_CLMUL; end else begin d_invalid_32bit = 1'b1; end
RV_CLMULH: if (EXTENSION_ZBC) begin d_aluop = ALUOP_CLMULH; end else begin d_invalid_32bit = 1'b1; end
RV_CLMULR: if (EXTENSION_ZBC) begin d_aluop = ALUOP_CLMULR; end else begin d_invalid_32bit = 1'b1; end
RV_BCLR: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BCLR; end else begin d_invalid_32bit = 1'b1; end
RV_BCLRI: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BCLR; d_rs2 = X0; d_imm = d_imm_i; d_alusrc_b = ALUSRCB_IMM; end else begin d_invalid_32bit = 1'b1; end
RV_BEXT: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BEXT; end else begin d_invalid_32bit = 1'b1; end
RV_BEXTI: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BEXT; d_rs2 = X0; d_imm = d_imm_i; d_alusrc_b = ALUSRCB_IMM; end else begin d_invalid_32bit = 1'b1; end
RV_BINV: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BINV; end else begin d_invalid_32bit = 1'b1; end
RV_BINVI: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BINV; d_rs2 = X0; d_imm = d_imm_i; d_alusrc_b = ALUSRCB_IMM; end else begin d_invalid_32bit = 1'b1; end
RV_BSET: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BSET; end else begin d_invalid_32bit = 1'b1; end
RV_BSETI: if (EXTENSION_ZBC) begin d_aluop = ALUOP_BSET; d_rs2 = X0; d_imm = d_imm_i; d_alusrc_b = ALUSRCB_IMM; end else begin d_invalid_32bit = 1'b1; end
RV_FENCE: begin d_rd = X0; end // NOP
RV_FENCE_I: begin d_invalid_32bit = DEBUG_SUPPORT && debug_mode; d_rd = X0; d_rs1 = X0; d_rs2 = X0; d_branchcond = BCOND_NZERO; d_imm[31] = 1'b1; end // FIXME this is probably busted now. Maybe implement as an exception?
RV_CSRRW: if (HAVE_CSR) begin d_imm = d_imm_i; d_csr_wen = 1'b1 ; d_csr_ren = |d_rd; d_csr_wtype = CSR_WTYPE_W; end else begin d_invalid_32bit = 1'b1; end

60
hdl/hazard3_ops.vh
View File

@ -1,26 +1,56 @@
// ALU operation selectors
localparam ALUOP_ADD = 4'h0;
localparam ALUOP_SUB = 4'h1;
localparam ALUOP_LT = 4'h2;
localparam ALUOP_LTU = 4'h4;
localparam ALUOP_AND = 4'h6;
localparam ALUOP_OR = 4'h7;
localparam ALUOP_XOR = 4'h8;
localparam ALUOP_SRL = 4'h9;
localparam ALUOP_SRA = 4'ha;
localparam ALUOP_SLL = 4'hb;
localparam ALUOP_MULDIV = 4'hc;
localparam ALUOP_ADD = 6'h00;
localparam ALUOP_SUB = 6'h01;
localparam ALUOP_LT = 6'h02;
localparam ALUOP_LTU = 6'h04;
localparam ALUOP_AND = 6'h06;
localparam ALUOP_OR = 6'h07;
localparam ALUOP_XOR = 6'h08;
localparam ALUOP_SRL = 6'h09;
localparam ALUOP_SRA = 6'h0a;
localparam ALUOP_SLL = 6'h0b;
localparam ALUOP_MULDIV = 6'h0c;
// Bitmanip ALU operations:
localparam ALUOP_SH1ADD = 6'h20;
localparam ALUOP_SH2ADD = 6'h21;
localparam ALUOP_SH3ADD = 6'h22;
localparam ALUOP_CLZ = 6'h23;
localparam ALUOP_CPOP = 6'h24;
localparam ALUOP_CTZ = 6'h25;
localparam ALUOP_ANDN = 6'h26; // Same LSBs as non-inverted
localparam ALUOP_ORN = 6'h27; // Same LSBs as non-inverted
localparam ALUOP_XNOR = 6'h28; // Same LSBs as non-inverted
localparam ALUOP_MAX = 6'h29;
localparam ALUOP_MAXU = 6'h2a;
localparam ALUOP_MIN = 6'h2b;
localparam ALUOP_MINU = 6'h2c;
localparam ALUOP_ORC_B = 6'h2d;
localparam ALUOP_REV8 = 6'h2e;
localparam ALUOP_ROL = 6'h2f;
localparam ALUOP_ROR = 6'h30;
localparam ALUOP_SEXT_B = 6'h31;
localparam ALUOP_SEXT_H = 6'h32;
localparam ALUOP_ZEXT_H = 6'h33;
localparam ALUOP_CLMUL = 6'h34;
localparam ALUOP_CLMULH = 6'h35;
localparam ALUOP_CLMULR = 6'h36;
localparam ALUOP_BCLR = 6'h37;
localparam ALUOP_BEXT = 6'h38;
localparam ALUOP_BINV = 6'h39;
localparam ALUOP_BSET = 6'h3a;
// Parameters to control ALU input muxes. Bypass mux paths are
// controlled by X, so D has no parameters to choose these.
localparam ALUSRCA_RS1 = 2'h0;
localparam ALUSRCA_PC = 2'h1;
localparam ALUSRCA_RS1 = 1'h0;
localparam ALUSRCA_PC = 1'h1;
localparam ALUSRCB_RS2 = 2'h0;
localparam ALUSRCB_IMM = 2'h1;
localparam ALUSRCB_RS2 = 1'h0;
localparam ALUSRCB_IMM = 1'h1;
localparam MEMOP_LW = 4'h0;
localparam MEMOP_LH = 4'h1;

5
hdl/hazard3_width_const.vh
View File

@ -5,10 +5,11 @@
parameter W_REGADDR = 5,
parameter W_ALUOP = 4,
parameter W_ALUSRC = 2,
parameter W_ALUOP = 6,
parameter W_ALUSRC = 1,
parameter W_MEMOP = 4,
parameter W_BCOND = 2,
parameter W_SHAMT = 5,
parameter W_EXCEPT = 4,
parameter W_MULOP = 3