重构:按功能拆分流水线模块、修复 BLT 指令、注释中文化

1) 模块拆分(便于单独仿真调试)
   把原来单文件的三模块结构按功能拆成 10 个独立文件,顶层只保留五级流水
   寄存器、子模块例化、PC/IF-ID/ID-EX 控制优先级与对外输出:
     cpu21_riscv_alu.v            组合 ALU(result2 用于 MUL 高位/余数)
     cpu21_bpb_8.v                8 项全相联分支目标缓冲
     cpu21_riscv_decoder.v        ID 级组合译码器
     cpu21_riscv_regfile.v        寄存器堆(写优先旁路)
     cpu21_riscv_forward_unit.v   EX 级前递网络
     cpu21_riscv_store_unit.v     EX 级存储数据对齐(sw/sb)
     cpu21_riscv_branch_unit.v    EX 级控制流裁决
     cpu21_riscv_hazard_unit.v    load-use 冒险与停顿/冲刷判定
     cpu21_riscv_irq_ctrl.v       中断优先级、ustatus/uepc、嵌套返回栈
     cpu21_riscv_perf_counters.v  性能计数器
   已把上述文件加入 Vivado 工程 sources_1;端口、时序与行为经 xsim 对同一
   ROM 逐周期回归验证,统计量完全一致。

2) 修复 BLT 指令(真值表第 28 行:opcode IR[6:2]=0x18、funct3=100、ALU_OP=SLT)
   此前 funct3=100 未译码,blt 被当作空指令执行,于是
   `blt s1,zero,loop` 的循环只执行一次就顺序落到退出代码并停机
   (现象:只输出第一个值后就不再运行)。
     - cpu21_riscv_decoder.v:新增 blt_o,OP_BRANCH 接受 funct3=100,
       并按真值表给出 ALU_SLT(beq/bne 仍为 SUB,bltu 仍为 SLTU)
     - cpu21_riscv_branch_unit.v:新增 blt_i,用有符号比较
       $signed(src1) < $signed(src2) 裁决
     - cpu21_riscv_redirect_int_bpb.v:新增 idex_blt_q 流水寄存器,
       IF 级 f_is_branch 纳入 funct3=100,使 blt 也参与 BPB 预测
   验证:blt 小程序正确输出 -3/-2/-1 后停机(cond_taken=2);
   benchmark 第 [4] 段完整输出 fffffff1..ffffffff;
   中断测试程序的中断入口与停留周期与改动前一致。

3) 注释中文化
   - cpu21_riscv_redirect_int_bpb.v 及全部新增子模块使用中文注释;
   - testbench/tb_cpu21_riscv_redirect_int_bpb.v 与新增的 tb_no_intr.v
     注释全部译为中文。
   - testbench/tb_no_intr.v:与中断测试平台同框架但不注入 irq 脉冲的对照
     测试平台(修复其 $dumpvars 引用了不存在的模块名,并改用 tb_no_intr.vcd
     避免与另一个 TB 的波形文件互相覆盖)。

4) 其他
   - testbench/tb_cpu21_riscv_redirect_int_bpb.v 的默认 ROM_FILE 改为
     cpu21_riscv_redirect_int_bpb_rom.hex(中断测试程序);
   - cpu21_riscv_redirect_int_bpb说明.md 增补"文件与模块划分"章节与 BLT 说明;
   - 真值表.txt 移动到 testbench/programs/ 下。
This commit is contained in:
2026-09-12 18:48:31 +08:00
parent abb2e161af
commit 266ec6a049
15 changed files with 1597 additions and 803 deletions
+248
View File
@@ -0,0 +1,248 @@
`timescale 1ns / 1ps
`default_nettype none
// -----------------------------------------------------------------------------
// ID 级组合译码器。
//
// 与原始 Logisim 控制器一致:操作码取标准 RISC-V 的 IR[6:2] 五位字段,
// 真值表中以十六进制形式存储该字段。
// 所有输出在无匹配时保持"无操作/不写回"的默认值。
// -----------------------------------------------------------------------------
module cpu21_riscv_decoder (
input wire [31:0] ir_i,
output reg [ 4:0] rs1_idx_o,
output reg [ 4:0] rs2_idx_o,
output reg [ 4:0] rd_o,
output reg [31:0] imm_o,
output reg [ 3:0] alu_op_o,
output reg [ 2:0] wb_sel_o,
output wire [11:0] csr_addr_o,
output reg uses_rs1_o,
output reg uses_rs2_o,
output reg reg_write_o,
output reg mem_to_reg_o,
output reg mem_write_o,
output reg mem_byte_o,
output reg alu_src_o,
output reg branch_o,
output reg beq_o,
output reg bne_o,
output reg blt_o,
output reg bltu_o,
output reg jal_o,
output reg jalr_o,
output reg ecall_o,
output reg uret_o,
output reg csr_set_o,
output reg csr_clear_o,
output reg csr_write_o
);
// CPU21 自定义操作码(取自所提供的真值表)。
localparam OP_LOAD = 5'h00;
localparam OP_R = 5'h0c;
localparam OP_I = 5'h04;
localparam OP_STORE = 5'h08;
localparam OP_JALR = 5'h19;
localparam OP_BRANCH = 5'h18;
localparam OP_JAL = 5'h1b;
localparam OP_SYS = 5'h1c;
localparam ALU_SLL = 4'd0;
localparam ALU_SRA = 4'd1;
localparam ALU_SRL = 4'd2;
localparam ALU_MUL = 4'd3;
localparam ALU_DIVU = 4'd4;
localparam ALU_ADD = 4'd5;
localparam ALU_SUB = 4'd6;
localparam ALU_AND = 4'd7;
localparam ALU_OR = 4'd8;
localparam ALU_XOR = 4'd9;
localparam ALU_SLT = 4'd11;
localparam ALU_SLTU = 4'd12;
wire [4:0] opcode = ir_i[6:2];
wire [2:0] funct3 = ir_i[14:12];
wire [6:0] funct7 = ir_i[31:25];
// CSR 地址字段在译码时直接旁路输出(供 EX 级读取 CSR)。
assign csr_addr_o = ir_i[31:20];
// 默认控制信号:全部为"无操作/不写回",再由下面的 case 覆盖。
always @* begin
rs1_idx_o = ir_i[19:15];
rs2_idx_o = ir_i[24:20];
rd_o = ir_i[11:7];
imm_o = {{20{ir_i[31]}}, ir_i[31:20]};
alu_op_o = ALU_ADD;
wb_sel_o = 3'd0;
uses_rs1_o = 1'b0;
uses_rs2_o = 1'b0;
reg_write_o = 1'b0;
mem_to_reg_o = 1'b0;
mem_write_o = 1'b0;
mem_byte_o = 1'b0;
alu_src_o = 1'b0;
branch_o = 1'b0;
beq_o = 1'b0;
bne_o = 1'b0;
blt_o = 1'b0;
bltu_o = 1'b0;
jal_o = 1'b0;
jalr_o = 1'b0;
ecall_o = 1'b0;
uret_o = 1'b0;
csr_set_o = 1'b0;
csr_clear_o = 1'b0;
csr_write_o = 1'b0;
case (opcode)
// R 型运算:由 funct3/funct7 决定具体操作。
OP_R: begin
uses_rs1_o = 1'b1;
uses_rs2_o = 1'b1;
reg_write_o = 1'b1;
// REMU R funct7=1/funct3=111
// ALU result2
if ((funct7 == 7'b0000001) && (funct3 == 3'b111)) begin
alu_op_o = ALU_DIVU;
wb_sel_o = 3'd3;
end else if ((funct7 == 7'b0000001) && (funct3 == 3'b000)) begin
alu_op_o = ALU_MUL;
end else begin
// 标准 R 型 funct3 译码;add/sub 与 sra/srl 由 funct7[5] 区分。
case (funct3)
3'b000: alu_op_o = (funct7[5] ? ALU_SUB : ALU_ADD);
3'b001: alu_op_o = ALU_SLL;
3'b010: alu_op_o = ALU_SLT;
3'b011: alu_op_o = ALU_SLTU;
3'b100: alu_op_o = ALU_XOR;
3'b101: alu_op_o = (funct7[5] ? ALU_SRA : ALU_SRL);
3'b110: alu_op_o = ALU_OR;
3'b111: alu_op_o = ALU_AND;
default: reg_write_o = 1'b0;
endcase
end
end
// I addi/slli/slti/xori/srai/srli/ori/andi
OP_I: begin
uses_rs1_o = 1'b1;
alu_src_o = 1'b1;
reg_write_o = 1'b1;
case (funct3)
3'b000: alu_op_o = ALU_ADD; // addi
3'b001: alu_op_o = ALU_SLL; // slli:立即数逻辑左移
3'b010: alu_op_o = ALU_SLT; // slti 1
3'b100: alu_op_o = ALU_XOR; // xori:立即数异或
3'b101:
alu_op_o = (funct7[5] ? ALU_SRA : ALU_SRL); // srai/srli/
3'b110: alu_op_o = ALU_OR; // ori:立即数或
3'b111: alu_op_o = ALU_AND; // andi
default: reg_write_o = 1'b0;
endcase
end
// 加载指令,当前仅支持 lwfunct3=010)。
OP_LOAD: begin
if (funct3 == 3'b010) begin
uses_rs1_o = 1'b1;
alu_src_o = 1'b1;
alu_op_o = ALU_ADD;
mem_to_reg_o = 1'b1;
reg_write_o = 1'b1;
wb_sel_o = 3'd1;
end
end
// 存储指令:swfunct3=010)与 sbfunct3=000)。
OP_STORE: begin
if ((funct3 == 3'b010) || (funct3 == 3'b000)) begin
uses_rs1_o = 1'b1;
uses_rs2_o = 1'b1;
alu_src_o = 1'b1;
alu_op_o = ALU_ADD;
mem_write_o = 1'b1;
mem_byte_o = (funct3 == 3'b000); // sb
imm_o = {{20{ir_i[31]}}, ir_i[31:25], ir_i[11:7]};
end
end
// beq000/bne001/blt100/bltu110
OP_BRANCH: begin
if ((funct3 == 3'b000) || (funct3 == 3'b001) ||
(funct3 == 3'b100) || (funct3 == 3'b110)) begin
uses_rs1_o = 1'b1;
uses_rs2_o = 1'b1;
branch_o = 1'b1;
beq_o = (funct3 == 3'b000);
bne_o = (funct3 == 3'b001);
blt_o = (funct3 == 3'b100);
bltu_o = (funct3 == 3'b110);
// 真值表:beq/bne 用减法比较,blt 用 ALU_SLTbltu 用 ALU_SLTU。
alu_op_o = blt_o ? ALU_SLT : (bltu_o ? ALU_SLTU : ALU_SUB);
imm_o = {{19{ir_i[31]}}, ir_i[31], ir_i[7], ir_i[30:25], ir_i[11:8], 1'b0};
end
end
// jal PC+4
OP_JAL: begin
jal_o = 1'b1;
reg_write_o = 1'b1;
wb_sel_o = 3'd2;
imm_o = {{11{ir_i[31]}}, ir_i[31], ir_i[19:12], ir_i[20], ir_i[30:21], 1'b0};
end
// jalr
OP_JALR: begin
if (funct3 == 3'b000) begin
jalr_o = 1'b1;
uses_rs1_o = 1'b1;
alu_src_o = 1'b1;
alu_op_o = ALU_ADD;
reg_write_o = 1'b1;
wb_sel_o = 3'd2;
end
end
// ecall / uret / CSR
OP_SYS: begin
// IR[21] URET ecall
if (funct3 == 3'b000) begin
if (ir_i[21]) begin
uret_o = 1'b1;
end else begin
ecall_o = 1'b1;
// 按文档说明,ecall 读取 a7(rs17) 与 a0(rs10)。
uses_rs1_o = 1'b1;
uses_rs2_o = 1'b1;
rs1_idx_o = 5'd17;
rs2_idx_o = 5'd10;
end
end else if (funct3 == 3'b001) begin
csr_write_o = 1'b1; // CSRRW:写 CSR,并把旧值写回 rd
uses_rs1_o = 1'b1;
alu_src_o = 1'b1;
imm_o = {27'b0, ir_i[19:15]};
reg_write_o = (rd_o != 5'd0);
wb_sel_o = 3'd4;
end else if (funct3 == 3'b110) begin
csr_set_o = 1'b1; // CSRRSI CSR
imm_o = {27'b0, ir_i[19:15]};
reg_write_o = (rd_o != 5'd0);
wb_sel_o = 3'd4;
end else if (funct3 == 3'b111) begin
csr_clear_o = 1'b1; // CSRRCI:清除 CSR 中的指定位
imm_o = {27'b0, ir_i[19:15]};
reg_write_o = (rd_o != 5'd0);
wb_sel_o = 3'd4;
end
end
default: begin
end
endcase
end
endmodule
`default_nettype wire