riscv/rocket-chip
1
0
Fork 0
Code Releases Activity

remove second RF write port

Browse Source 
load miss writebacks are treated like mul/div now.
This commit is contained in:
Andrew Waterman 2012-01-02 02:51:30 -08:00
parent ffe23a1ee8
commit 3045b33460
6 changed files with 63 additions and 66 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
rocket/src/main/scala/cpu.scala
View File

@ -101,9 +101,11 @@ class rocketProc extends Component
arb.io.cpu.req_tag := dpath.io.dmem.req_tag; arb.io.cpu.req_tag := dpath.io.dmem.req_tag;
ctrl.io.dmem.req_rdy := dtlb.io.cpu.req_rdy && arb.io.cpu.req_rdy; ctrl.io.dmem.req_rdy := dtlb.io.cpu.req_rdy && arb.io.cpu.req_rdy;
ctrl.io.dmem.resp_miss := arb.io.cpu.resp_miss; ctrl.io.dmem.resp_miss := arb.io.cpu.resp_miss;
ctrl.io.dmem.resp_replay:= arb.io.cpu.resp_replay;
ctrl.io.dmem.resp_nack := arb.io.cpu.resp_nack; ctrl.io.dmem.resp_nack := arb.io.cpu.resp_nack;
dpath.io.dmem.resp_val := arb.io.cpu.resp_val; dpath.io.dmem.resp_val := arb.io.cpu.resp_val;
dpath.io.dmem.resp_replay := io.dmem.resp_replay; dpath.io.dmem.resp_miss := arb.io.cpu.resp_miss;
dpath.io.dmem.resp_replay := arb.io.cpu.resp_replay;
dpath.io.dmem.resp_tag := arb.io.cpu.resp_tag; dpath.io.dmem.resp_tag := arb.io.cpu.resp_tag;
dpath.io.dmem.resp_data := arb.io.cpu.resp_data; dpath.io.dmem.resp_data := arb.io.cpu.resp_data;
dpath.io.dmem.resp_data_subword := io.dmem.resp_data_subword; dpath.io.dmem.resp_data_subword := io.dmem.resp_data_subword;

56
rocket/src/main/scala/ctrl.scala
View File

@ -64,10 +64,8 @@ class ioCtrlDpath extends Bundle()
val mem_waddr = UFix(5,'input); // write addr from memory stage val mem_waddr = UFix(5,'input); // write addr from memory stage
val wb_waddr = UFix(5,'input); // write addr from writeback stage val wb_waddr = UFix(5,'input); // write addr from writeback stage
val status = Bits(17, 'input); val status = Bits(17, 'input);
val sboard_clr0 = Bool('input); val sboard_clr = Bool('input);
val sboard_clr0a = UFix(5, 'input); val sboard_clra = UFix(5, 'input);
val sboard_clr1 = Bool('input);
val sboard_clr1a = UFix(5, 'input);
val mem_valid = Bool('input); // high if there's a valid (not flushed) instruction in mem stage val mem_valid = Bool('input); // high if there's a valid (not flushed) instruction in mem stage
val irq_timer = Bool('input); val irq_timer = Bool('input);
val irq_ipi = Bool('input); val irq_ipi = Bool('input);
@ -78,7 +76,7 @@ class ioCtrlAll extends Bundle()
val dpath = new ioCtrlDpath(); val dpath = new ioCtrlDpath();
val console = new ioConsole(List("rdy")); val console = new ioConsole(List("rdy"));
val imem = new ioImem(List("req_val", "req_rdy", "resp_val")).flip(); val imem = new ioImem(List("req_val", "req_rdy", "resp_val")).flip();
val dmem = new ioDmem(List("req_val", "req_kill", "req_rdy", "req_cmd", "req_type", "resp_miss", "resp_nack")).flip(); val dmem = new ioDmem(List("req_val", "req_kill", "req_rdy", "req_cmd", "req_type", "resp_miss", "resp_replay", "resp_nack")).flip();
val host = new ioHost(List("start")); val host = new ioHost(List("start"));
val dtlb_val = Bool('output) val dtlb_val = Bool('output)
val dtlb_rdy = Bool('input); val dtlb_rdy = Bool('input);
@ -186,26 +184,26 @@ class rocketCtrl extends Component
JALR_R-> List(Y, BR_JR, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_PC, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), JALR_R-> List(Y, BR_JR, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_PC, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
RDNPC-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_PC, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), RDNPC-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_PC, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LB-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_B, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LB-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_B, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LH-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_H, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LH-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_H, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LW-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_W, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LW-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_W, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LD-> List(xpr64, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_D, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LD-> List(xpr64, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_D, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LBU-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_BU,N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LBU-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_BU,N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LHU-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_HU,N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LHU-> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_HU,N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LWU-> List(xpr64, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_WU,N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LWU-> List(xpr64, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XRD, MT_WU,N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
SB-> List(Y, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_B, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), SB-> List(Y, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_B, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
SH-> List(Y, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_H, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), SH-> List(Y, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_H, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
SW-> List(Y, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_W, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), SW-> List(Y, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_W, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
SD-> List(xpr64, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_D, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), SD-> List(xpr64, BR_N, REN_Y,REN_Y,A2_SPLIT,A1_RS1,DW_XPR,FN_ADD, M_Y,M_XWR, MT_D, N,MUL_X, N,DIV_X, WEN_N,WA_X, WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOADD_W-> List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_ADD, MT_W, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOADD_W-> List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_ADD, MT_W, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOSWAP_W->List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_SWAP,MT_W, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOSWAP_W->List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_SWAP,MT_W, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOAND_W-> List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_AND, MT_W, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOAND_W-> List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_AND, MT_W, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOOR_W-> List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_OR, MT_W, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOOR_W-> List(Y, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_OR, MT_W, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOADD_D-> List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_ADD, MT_D, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOADD_D-> List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_ADD, MT_D, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOSWAP_D->List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_SWAP,MT_D, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOSWAP_D->List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_SWAP,MT_D, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOAND_D-> List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_AND, MT_D, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOAND_D-> List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_AND, MT_D, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
AMOOR_D-> List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_OR, MT_D, N,MUL_X, N,DIV_X, WEN_N,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N), AMOOR_D-> List(xpr64, BR_N, REN_Y,REN_Y,A2_0, A1_RS1,DW_XPR,FN_ADD, M_Y,M_XA_OR, MT_D, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_X, REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
LUI-> List(Y, BR_N, REN_N,REN_Y,A2_0, A1_LUI,DW_XPR,FN_ADD, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_ALU,REN_N,WEN_N,I_X ,SYNC_N,N,N,N), LUI-> List(Y, BR_N, REN_N,REN_Y,A2_0, A1_LUI,DW_XPR,FN_ADD, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_ALU,REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
SLTI -> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_SLT, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_ALU,REN_N,WEN_N,I_X ,SYNC_N,N,N,N), SLTI -> List(Y, BR_N, REN_N,REN_Y,A2_SEXT, A1_RS1,DW_XPR,FN_SLT, M_N,M_X, MT_X, N,MUL_X, N,DIV_X, WEN_Y,WA_RD,WB_ALU,REN_N,WEN_N,I_X ,SYNC_N,N,N,N),
@ -318,10 +316,8 @@ class rocketCtrl extends Component
sboard.io.set := wb_reg_div_mul_val | dcache_miss; sboard.io.set := wb_reg_div_mul_val | dcache_miss;
sboard.io.seta := io.dpath.wb_waddr; sboard.io.seta := io.dpath.wb_waddr;
sboard.io.clr0 := io.dpath.sboard_clr0; sboard.io.clr := io.dpath.sboard_clr;
sboard.io.clr0a := io.dpath.sboard_clr0a; sboard.io.clra := io.dpath.sboard_clra;
sboard.io.clr1 := io.dpath.sboard_clr1;
sboard.io.clr1a := io.dpath.sboard_clr1a;
val id_stall_raddr2 = sboard.io.stalla; val id_stall_raddr2 = sboard.io.stalla;
val id_stall_raddr1 = sboard.io.stallb; val id_stall_raddr1 = sboard.io.stallb;
@ -613,6 +609,11 @@ class rocketCtrl extends Component
(id_wen.toBool && (id_waddr === io.dpath.wb_waddr))); (id_wen.toBool && (id_waddr === io.dpath.wb_waddr)));
val data_hazard = data_hazard_ex || data_hazard_mem || data_hazard_wb; val data_hazard = data_hazard_ex || data_hazard_mem || data_hazard_wb;
// for divider, multiplier, load miss writeback
val mem_wb = Reg(io.dmem.resp_replay, resetVal = Bool(false)) // delayed for subword extension
val mul_wb = io.dpath.mul_result_val && !io.dmem.resp_replay;
val div_wb = io.dpath.div_result_val && !io.dpath.mul_result_val && !io.dmem.resp_replay;
val ctrl_stalld = val ctrl_stalld =
!take_pc && !take_pc &&
@ -627,15 +628,12 @@ class rocketCtrl extends Component
id_div_val.toBool && !io.dpath.div_rdy || id_div_val.toBool && !io.dpath.div_rdy ||
id_mul_val.toBool && !io.dpath.mul_rdy || id_mul_val.toBool && !io.dpath.mul_rdy ||
io.dpath.div_result_val || io.dpath.div_result_val ||
io.dpath.mul_result_val io.dpath.mul_result_val ||
mem_wb
); );
val ctrl_killd = take_pc || ctrl_stalld; val ctrl_killd = take_pc || ctrl_stalld;
val ctrl_killf = take_pc || !io.imem.resp_val; val ctrl_killf = take_pc || !io.imem.resp_val;
// for divider, multiplier writeback
val mul_wb = io.dpath.mul_result_val;
val div_wb = io.dpath.div_result_val & !mul_wb;
io.flush_inst := mem_reg_flush_inst; io.flush_inst := mem_reg_flush_inst;

13
rocket/src/main/scala/ctrl_util.scala
View File

@ -6,10 +6,8 @@ import Constants._;
class ioCtrlSboard extends Bundle() class ioCtrlSboard extends Bundle()
{ {
val clr0 = Bool('input); val clr = Bool('input);
val clr0a = UFix(5, 'input); val clra = UFix(5, 'input);
val clr1 = Bool('input);
val clr1a = UFix(5, 'input);
val set = Bool('input); val set = Bool('input);
val seta = UFix(5, 'input); val seta = UFix(5, 'input);
val raddra = UFix(5, 'input); val raddra = UFix(5, 'input);
@ -25,10 +23,9 @@ class rocketCtrlSboard extends Component
override val io = new ioCtrlSboard(); override val io = new ioCtrlSboard();
val reg_busy = Reg(width = 32, resetVal = Bits(0, 32)); val reg_busy = Reg(width = 32, resetVal = Bits(0, 32));
val set_mask = Mux(io.set, UFix(1,1) << io.seta, UFix(0,32)); val set_mask = io.set.toUFix << io.seta;
val clr0_mask = Mux(io.clr0, ~(UFix(1,1) << io.clr0a), ~UFix(0,32)); val clr_mask = ~(io.clr.toUFix << io.clra);
val clr1_mask = Mux(io.clr1, ~(UFix(1,1) << io.clr1a), ~UFix(0,32)); reg_busy <== (reg_busy | set_mask) & clr_mask
reg_busy <== ((reg_busy | set_mask) & clr0_mask) & clr1_mask;
io.stalla := reg_busy(io.raddra).toBool; io.stalla := reg_busy(io.raddra).toBool;
io.stallb := reg_busy(io.raddrb).toBool; io.stallb := reg_busy(io.raddrb).toBool;

49
rocket/src/main/scala/dpath.scala
View File

@ -11,6 +11,7 @@ class ioDpathDmem extends Bundle()
val req_tag = UFix(CPU_TAG_BITS, 'output); val req_tag = UFix(CPU_TAG_BITS, 'output);
val req_data = Bits(64, 'output); val req_data = Bits(64, 'output);
val resp_val = Bool('input); val resp_val = Bool('input);
val resp_miss = Bool('input);
val resp_replay = Bool('input); val resp_replay = Bool('input);
val resp_tag = Bits(CPU_TAG_BITS, 'input); val resp_tag = Bits(CPU_TAG_BITS, 'input);
val resp_data = Bits(64, 'input); val resp_data = Bits(64, 'input);
@ -78,7 +79,6 @@ class rocketDpath extends Component
val ex_reg_pc_plus4 = Reg() { UFix() }; val ex_reg_pc_plus4 = Reg() { UFix() };
val ex_reg_inst = Reg() { Bits() }; val ex_reg_inst = Reg() { Bits() };
val ex_reg_raddr2 = Reg() { UFix() }; val ex_reg_raddr2 = Reg() { UFix() };
val ex_reg_raddr1 = Reg() { UFix() };
val ex_reg_rs2 = Reg() { Bits() }; val ex_reg_rs2 = Reg() { Bits() };
val ex_reg_rs1 = Reg() { Bits() }; val ex_reg_rs1 = Reg() { Bits() };
val ex_reg_waddr = Reg() { UFix() }; val ex_reg_waddr = Reg() { UFix() };
@ -115,6 +115,7 @@ class rocketDpath extends Component
val wb_reg_ctrl_wen = Reg(resetVal = Bool(false)); val wb_reg_ctrl_wen = Reg(resetVal = Bool(false));
val r_dmem_resp_val = Reg(resetVal = Bool(false)); val r_dmem_resp_val = Reg(resetVal = Bool(false));
val r_dmem_resp_replay = Reg(resetVal = Bool(false));
val r_dmem_resp_waddr = Reg() { UFix() }; val r_dmem_resp_waddr = Reg() { UFix() };
// instruction fetch stage // instruction fetch stage
@ -131,7 +132,8 @@ class rocketDpath extends Component
val ex_branch_target = ex_reg_pc + branch_adder_rhs.toUFix; val ex_branch_target = ex_reg_pc + branch_adder_rhs.toUFix;
btb.io.correct_target := Mux(io.ctrl.ex_jr, ex_jr_target, ex_branch_target); val jr_br_target = Mux(io.ctrl.ex_jr, ex_jr_target, ex_branch_target);
btb.io.correct_target := jr_br_target
val if_next_pc = val if_next_pc =
Mux(io.ctrl.sel_pc === PC_BTB, if_btb_target, Mux(io.ctrl.sel_pc === PC_BTB, if_btb_target,
@ -158,7 +160,7 @@ class rocketDpath extends Component
btb.io.hit ^^ io.ctrl.btb_hit; btb.io.hit ^^ io.ctrl.btb_hit;
btb.io.wen ^^ io.ctrl.wen_btb; btb.io.wen ^^ io.ctrl.wen_btb;
btb.io.correct_pc4 := ex_reg_pc_plus4; btb.io.correct_pc4 := ex_reg_pc_plus4;
io.ctrl.btb_match := id_reg_pc === btb.io.correct_target; io.ctrl.btb_match := id_reg_pc === jr_br_target;
// instruction decode stage // instruction decode stage
when (!io.ctrl.stalld) { when (!io.ctrl.stalld) {
@ -187,32 +189,35 @@ class rocketDpath extends Component
val id_rdata1 = rfile.io.r1.data; val id_rdata1 = rfile.io.r1.data;
// destination register selection // destination register selection
val id_ctrl_ll_wb = io.ctrl.div_wb || io.ctrl.mul_wb || r_dmem_resp_replay;
val id_waddr = val id_waddr =
Mux(io.ctrl.div_wb, div_result_tag, Mux(r_dmem_resp_replay, r_dmem_resp_waddr,
Mux(io.ctrl.mul_wb, mul_result_tag, Mux(io.ctrl.mul_wb, mul_result_tag,
Mux(io.ctrl.div_wb, div_result_tag,
Mux(io.ctrl.sel_wa === WA_RD, id_reg_inst(31,27).toUFix, Mux(io.ctrl.sel_wa === WA_RD, id_reg_inst(31,27).toUFix,
Mux(io.ctrl.sel_wa === WA_RA, RA, Mux(io.ctrl.sel_wa === WA_RA, RA,
UFix(0, 5))))); UFix(0, 5))))));
// bypass muxes // bypass muxes
val rs1_mem_lu_bypass = id_raddr1 != UFix(0, 5) && io.ctrl.mem_load && id_raddr1 === mem_reg_waddr; val rs1_mem_lu_bypass = id_raddr1 != UFix(0, 5) && io.ctrl.mem_load && id_raddr1 === mem_reg_waddr;
val id_rs1 = val id_rs1 =
Mux(r_dmem_resp_replay, io.dmem.resp_data_subword,
Mux(io.ctrl.div_wb, div_result, Mux(io.ctrl.div_wb, div_result,
Mux(io.ctrl.mul_wb, mul_result, Mux(io.ctrl.mul_wb, mul_result,
Mux(id_raddr1 != UFix(0, 5) && ex_reg_ctrl_wen && id_raddr1 === ex_reg_waddr, ex_wdata, Mux(id_raddr1 != UFix(0, 5) && (ex_reg_ctrl_wen || ex_reg_ctrl_ll_wb) && id_raddr1 === ex_reg_waddr, ex_wdata,
Mux(id_raddr1 != UFix(0, 5) && mem_reg_ctrl_wen && id_raddr1 === mem_reg_waddr, mem_reg_wdata,
Mux(rs1_mem_lu_bypass, io.dmem.resp_data, Mux(rs1_mem_lu_bypass, io.dmem.resp_data,
Mux(id_raddr1 != UFix(0, 5) && (mem_reg_ctrl_wen || mem_reg_ctrl_ll_wb) && id_raddr1 === mem_reg_waddr, mem_reg_wdata,
Mux(id_raddr1 != UFix(0, 5) && r_dmem_resp_val && id_raddr1 === r_dmem_resp_waddr, io.dmem.resp_data_subword, Mux(id_raddr1 != UFix(0, 5) && r_dmem_resp_val && id_raddr1 === r_dmem_resp_waddr, io.dmem.resp_data_subword,
Mux(id_raddr1 != UFix(0, 5) && wb_reg_ctrl_wen && id_raddr1 === wb_reg_waddr, wb_reg_wdata, Mux(id_raddr1 != UFix(0, 5) && (wb_reg_ctrl_wen || wb_reg_ctrl_ll_wb) && id_raddr1 === wb_reg_waddr, wb_reg_wdata,
id_rdata1))))))); id_rdata1))))))));
val rs2_mem_lu_bypass = id_raddr2 != UFix(0, 5) && io.ctrl.mem_load && id_raddr2 === mem_reg_waddr; val rs2_mem_lu_bypass = id_raddr2 != UFix(0, 5) && io.ctrl.mem_load && id_raddr2 === mem_reg_waddr;
val id_rs2 = val id_rs2 =
Mux(id_raddr2 != UFix(0, 5) && ex_reg_ctrl_wen && id_raddr2 === ex_reg_waddr, ex_wdata, Mux(id_raddr2 != UFix(0, 5) && (ex_reg_ctrl_wen || ex_reg_ctrl_ll_wb) && id_raddr2 === ex_reg_waddr, ex_wdata,
Mux(id_raddr2 != UFix(0, 5) && mem_reg_ctrl_wen && id_raddr2 === mem_reg_waddr, mem_reg_wdata,
Mux(rs2_mem_lu_bypass, io.dmem.resp_data, Mux(rs2_mem_lu_bypass, io.dmem.resp_data,
Mux(id_raddr2 != UFix(0, 5) && (mem_reg_ctrl_wen || mem_reg_ctrl_ll_wb) && id_raddr2 === mem_reg_waddr, mem_reg_wdata,
Mux(id_raddr2 != UFix(0, 5) && r_dmem_resp_val && id_raddr2 === r_dmem_resp_waddr, io.dmem.resp_data_subword, Mux(id_raddr2 != UFix(0, 5) && r_dmem_resp_val && id_raddr2 === r_dmem_resp_waddr, io.dmem.resp_data_subword,
Mux(id_raddr2 != UFix(0, 5) && wb_reg_ctrl_wen && id_raddr2 === wb_reg_waddr, wb_reg_wdata, Mux(id_raddr2 != UFix(0, 5) && (wb_reg_ctrl_wen || wb_reg_ctrl_ll_wb) && id_raddr2 === wb_reg_waddr, wb_reg_wdata,
id_rdata2))))); id_rdata2)))));
io.ctrl.mem_lu_bypass := rs1_mem_lu_bypass || rs2_mem_lu_bypass; io.ctrl.mem_lu_bypass := rs1_mem_lu_bypass || rs2_mem_lu_bypass;
@ -223,7 +228,6 @@ class rocketDpath extends Component
ex_reg_pc_plus4 <== id_reg_pc_plus4; ex_reg_pc_plus4 <== id_reg_pc_plus4;
ex_reg_inst <== id_reg_inst; ex_reg_inst <== id_reg_inst;
ex_reg_raddr2 <== id_raddr2; ex_reg_raddr2 <== id_raddr2;
ex_reg_raddr1 <== id_raddr1;
ex_reg_rs2 <== id_rs2; ex_reg_rs2 <== id_rs2;
ex_reg_rs1 <== id_rs1; ex_reg_rs1 <== id_rs1;
ex_reg_waddr <== id_waddr; ex_reg_waddr <== id_waddr;
@ -233,7 +237,7 @@ class rocketDpath extends Component
ex_reg_ctrl_fn_alu <== io.ctrl.fn_alu; ex_reg_ctrl_fn_alu <== io.ctrl.fn_alu;
ex_reg_ctrl_mul_fn <== io.ctrl.mul_fn; ex_reg_ctrl_mul_fn <== io.ctrl.mul_fn;
ex_reg_ctrl_div_fn <== io.ctrl.div_fn; ex_reg_ctrl_div_fn <== io.ctrl.div_fn;
ex_reg_ctrl_ll_wb <== io.ctrl.div_wb | io.ctrl.mul_wb; // TODO: verify ex_reg_ctrl_ll_wb <== id_ctrl_ll_wb;
ex_reg_ctrl_sel_wb <== io.ctrl.sel_wb; ex_reg_ctrl_sel_wb <== io.ctrl.sel_wb;
ex_reg_ctrl_ren_pcr <== io.ctrl.ren_pcr; ex_reg_ctrl_ren_pcr <== io.ctrl.ren_pcr;
@ -371,6 +375,7 @@ class rocketDpath extends Component
// writeback stage // writeback stage
r_dmem_resp_val <== io.dmem.resp_val; r_dmem_resp_val <== io.dmem.resp_val;
r_dmem_resp_replay <== io.dmem.resp_replay;
r_dmem_resp_waddr <== io.dmem.resp_tag.toUFix r_dmem_resp_waddr <== io.dmem.resp_tag.toUFix
wb_reg_waddr <== mem_reg_waddr; wb_reg_waddr <== mem_reg_waddr;
@ -381,27 +386,21 @@ class rocketDpath extends Component
wb_reg_ctrl_wen <== Bool(false); wb_reg_ctrl_wen <== Bool(false);
} }
otherwise { otherwise {
wb_reg_ctrl_wen <== mem_reg_ctrl_wen; wb_reg_ctrl_wen <== mem_reg_ctrl_wen && !io.dmem.resp_miss;
} }
// crossbar/sign extension for 8/16 bit loads (moved to earlier in file) // crossbar/sign extension for 8/16 bit loads (moved to earlier in file)
// regfile write // regfile write
rfile.io.w0.addr := wb_reg_waddr; rfile.io.w0.addr := wb_reg_waddr;
rfile.io.w0.en := wb_reg_ctrl_wen | wb_reg_ctrl_ll_wb; rfile.io.w0.en := wb_reg_ctrl_wen || wb_reg_ctrl_ll_wb;
rfile.io.w0.data := wb_reg_wdata; rfile.io.w0.data := Mux(Reg(io.ctrl.mem_load), io.dmem.resp_data_subword, wb_reg_wdata);
rfile.io.w1.addr := r_dmem_resp_waddr;
rfile.io.w1.en := r_dmem_resp_val;
rfile.io.w1.data := io.dmem.resp_data_subword;
io.ctrl.wb_waddr := wb_reg_waddr; io.ctrl.wb_waddr := wb_reg_waddr;
// scoreboard clear (for div/mul and D$ load miss writebacks) // scoreboard clear (for div/mul and D$ load miss writebacks)
io.ctrl.sboard_clr0 := wb_reg_ctrl_ll_wb; io.ctrl.sboard_clr := id_ctrl_ll_wb;
io.ctrl.sboard_clr0a := wb_reg_waddr; io.ctrl.sboard_clra := id_waddr;
io.ctrl.sboard_clr1 := r_dmem_resp_val;
io.ctrl.sboard_clr1a := r_dmem_resp_waddr;
// processor control regfile write // processor control regfile write
pcr.io.w.addr := mem_reg_raddr2; pcr.io.w.addr := mem_reg_raddr2;

2
rocket/src/main/scala/dpath_util.scala
View File

@ -219,14 +219,12 @@ class ioRegfile extends Bundle()
val r0 = new ioReadPort(); val r0 = new ioReadPort();
val r1 = new ioReadPort(); val r1 = new ioReadPort();
val w0 = new ioWritePort(); val w0 = new ioWritePort();
val w1 = new ioWritePort();
} }
class rocketDpathRegfile extends Component class rocketDpathRegfile extends Component
{ {
override val io = new ioRegfile(); override val io = new ioRegfile();
val regfile = Mem(32, io.w0.en && (io.w0.addr != UFix(0,5)), io.w0.addr, io.w0.data); val regfile = Mem(32, io.w0.en && (io.w0.addr != UFix(0,5)), io.w0.addr, io.w0.data);
regfile.write(io.w1.en && (io.w1.addr != UFix(0,5)), io.w1.addr, io.w1.data);
io.r0.data := Mux((io.r0.addr === UFix(0, 5)) || !io.r0.en, Bits(0, 64), regfile(io.r0.addr)); io.r0.data := Mux((io.r0.addr === UFix(0, 5)) || !io.r0.en, Bits(0, 64), regfile(io.r0.addr));
io.r1.data := Mux((io.r1.addr === UFix(0, 5)) || !io.r1.en, Bits(0, 64), regfile(io.r1.addr)); io.r1.data := Mux((io.r1.addr === UFix(0, 5)) || !io.r1.en, Bits(0, 64), regfile(io.r1.addr));
} }

5
rocket/src/main/scala/ptw.scala
View File

@ -7,7 +7,7 @@ import scala.math._;
class ioDmemArbiter extends Bundle class ioDmemArbiter extends Bundle
{ {
val ptw = new ioDmem(List("req_val", "req_rdy", "req_cmd", "req_type", "req_idx", "req_ppn", "resp_data", "resp_val", "resp_nack")); val ptw = new ioDmem(List("req_val", "req_rdy", "req_cmd", "req_type", "req_idx", "req_ppn", "resp_data", "resp_val", "resp_replay", "resp_nack"));
val cpu = new ioDmem(); val cpu = new ioDmem();
val mem = new ioDmem().flip(); val mem = new ioDmem().flip();
} }
@ -40,6 +40,9 @@ class rocketDmemArbiter extends Component
io.cpu.resp_val := io.mem.resp_val && !io.mem.resp_tag(0).toBool; io.cpu.resp_val := io.mem.resp_val && !io.mem.resp_tag(0).toBool;
io.ptw.resp_val := io.mem.resp_val && io.mem.resp_tag(0).toBool; io.ptw.resp_val := io.mem.resp_val && io.mem.resp_tag(0).toBool;
io.cpu.resp_replay := io.mem.resp_replay && !io.mem.resp_tag(0).toBool;
io.ptw.resp_replay := io.mem.resp_replay && io.mem.resp_tag(0).toBool;
io.ptw.resp_data := io.mem.resp_data; io.ptw.resp_data := io.mem.resp_data;
io.cpu.resp_data := io.mem.resp_data; io.cpu.resp_data := io.mem.resp_data;
io.cpu.resp_tag := io.mem.resp_tag >> UFix(1); io.cpu.resp_tag := io.mem.resp_tag >> UFix(1);