Optimization. Able to remove hptw address muxes from the E stage.

pipelined/src/cache/cache.sv

@@ -52,6 +52,7 @@ module cache #(parameter LINELEN,  NUMLINES,  NUMWAYS, LOGBWPL, WORDLEN, MUXINTE
   input logic                   IgnoreRequestTLB,
   input logic                   TrapM, 
   input logic                   Cacheable,
+  input logic                   SelReplay,
    // Bus fsm interface
   output logic                  CacheFetchLine,
   output logic                  CacheWriteLine,
@@ -123,7 +124,7 @@ module cache #(parameter LINELEN,  NUMLINES,  NUMWAYS, LOGBWPL, WORDLEN, MUXINTE
   // and FlushAdr when handling D$ flushes  
   mux3 #(SETLEN) AdrSelMux(
     .d0(NextAdr[SETTOP-1:OFFSETLEN]), .d1(PAdr[SETTOP-1:OFFSETLEN]), .d2(FlushAdr),
-    .s({SelFlush, SelAdr}), .y(RAdr));
+    .s({SelFlush, (SelAdr | SelReplay)}), .y(RAdr));
 
   // Array of cache ways, along with victim, hit, dirty, and read merging logic
   cacheway #(NUMLINES, LINELEN, TAGLEN, OFFSETLEN, SETLEN) 

pipelined/src/ifu/ifu.sv

@@ -222,6 +222,7 @@ module ifu (
              .CacheFetchLine(ICacheFetchLine),
              .CacheWriteLine(), .ReadDataWord(FinalInstrRawF),
              .Cacheable(CacheableF),
+             .SelReplay('0),
              .CacheMiss(ICacheMiss), .CacheAccess(ICacheAccess),
              .ByteMask('0), .WordCount('0), .SelBusWord('0),
              .FinalWriteData('0),

pipelined/src/lsu/lsu.sv

@@ -104,7 +104,6 @@ module lsu (
   logic [6:0]               LSUFunct7M;
   logic [1:0]               LSUAtomicM;
   (* mark_debug = "true" *)  logic [`XLEN+1:0] 		   PreLSUPAdrM;
-  logic [11:0]              LSUAdrE;  
   logic                     SelDTIM;
   logic                     CPUBusy;
   logic                     DCacheStallM;
@@ -118,7 +117,8 @@ module lsu (
   logic [`LLEN-1:0]         IMAFWriteDataM;
   logic [`LLEN-1:0]         ReadDataM;
   logic [(`LLEN-1)/8:0]     ByteMaskM;
-  
+  logic                     SelReplay;
+    
   flopenrc #(`XLEN) AddressMReg(clk, reset, FlushM, ~StallM, IEUAdrE, IEUAdrM);
   assign IEUAdrExtM = {2'b00, IEUAdrM}; 
   assign IEUAdrExtE = {2'b00, IEUAdrE}; 
@@ -131,17 +131,16 @@ module lsu (
 
   if(`VIRTMEM_SUPPORTED) begin : VIRTMEM_SUPPORTED
     lsuvirtmem lsuvirtmem(.clk, .reset, .StallW, .MemRWM(NonDTIMMemRWM), .AtomicM, .ITLBMissF, .ITLBWriteF,
-      .DTLBMissM, .DTLBWriteM, .InstrDAPageFaultF, .DataDAPageFaultM, 
+      .DTLBMissM, .DTLBWriteM, .InstrDAPageFaultF, .DataDAPageFaultM, .SelReplay,
       .TrapM, .DCacheStallM, .SATP_REGW, .PCF,
       .STATUS_MXR, .STATUS_SUM, .STATUS_MPRV, .STATUS_MPP, .PrivilegeModeW,
       .ReadDataM(ReadDataM[`XLEN-1:0]), .WriteDataM, .Funct3M, .LSUFunct3M, .Funct7M, .LSUFunct7M,
-      .IEUAdrExtM, .PTE, .IMWriteDataM, .PageType, .PreLSURWM, .LSUAtomicM, .IEUAdrE,
+      .IEUAdrExtM, .PTE, .IMWriteDataM, .PageType, .PreLSURWM, .LSUAtomicM,
-      .LSUAdrE, .PreLSUPAdrM, .CPUBusy, .InterlockStall, .SelHPTW,
+      .PreLSUPAdrM, .CPUBusy, .InterlockStall, .SelHPTW,
       .IgnoreRequestTLB);
   end else begin
     assign {InterlockStall, SelHPTW, PTE, PageType, DTLBWriteM, ITLBWriteF, IgnoreRequestTLB} = '0;
     assign CPUBusy = StallW; assign PreLSURWM = NonDTIMMemRWM; 
-    assign LSUAdrE = IEUAdrE[11:0]; 
     assign PreLSUPAdrM = IEUAdrExtM;
     assign LSUFunct3M = Funct3M;  assign LSUFunct7M = Funct7M; assign LSUAtomicM = AtomicM;
     assign IMWriteDataM = WriteDataM;
@@ -251,9 +250,9 @@ module lsu (
       cache #(.LINELEN(`DCACHE_LINELENINBITS), .NUMLINES(`DCACHE_WAYSIZEINBYTES*8/LINELEN),
               .NUMWAYS(`DCACHE_NUMWAYS), .LOGBWPL(LOGBWPL), .WORDLEN(`LLEN), .MUXINTERVAL(`XLEN), .DCACHE(1)) dcache(
         .clk, .reset, .CPUBusy, .SelBusWord, .RW(LSURWM), .Atomic(LSUAtomicM),
-        .FlushCache(FlushDCacheM), .NextAdr(LSUAdrE), .PAdr(LSUPAdrM), 
+        .FlushCache(FlushDCacheM), .NextAdr(IEUAdrE[11:0]), .PAdr(LSUPAdrM), 
         .ByteMask(ByteMaskM), .WordCount,
-        .FinalWriteData(LSUWriteDataM), .Cacheable(CacheableM),
+        .FinalWriteData(LSUWriteDataM), .Cacheable(CacheableM), .SelReplay,
         .CacheStall(DCacheStallM), .CacheMiss(DCacheMiss), .CacheAccess(DCacheAccess),
         .IgnoreRequestTLB, .TrapM, .CacheCommitted(DCacheCommittedM), 
         .CacheBusAdr(DCacheBusAdr), .ReadDataWord(ReadDataWordM), 

pipelined/src/lsu/lsuvirtmen.sv

@@ -38,6 +38,7 @@ module lsuvirtmem(
   input logic                 DTLBMissM,
   output logic                DTLBWriteM,
   input logic                 InstrDAPageFaultF,
+  output logic                SelReplay,
   input logic                 DataDAPageFaultM,
   input logic                 TrapM,
   input logic                 DCacheStallM,
@@ -52,13 +53,11 @@ module lsuvirtmem(
   output logic [2:0]          LSUFunct3M,
   input logic [6:0]           Funct7M,
   output logic [6:0]          LSUFunct7M,
-  input logic [`XLEN-1:0]     IEUAdrE,
   output logic [`XLEN-1:0]    PTE,
   output logic [`XLEN-1:0]    IMWriteDataM,
   output logic [1:0]          PageType,
   output logic [1:0]          PreLSURWM,
   output logic [1:0]          LSUAtomicM,
-  output logic [11:0]         LSUAdrE,
   output logic [`XLEN+1:0] PreLSUPAdrM,
   input logic [`XLEN+1:0]     IEUAdrExtM, // *** can move internally.
                   
@@ -74,7 +73,6 @@ module lsuvirtmem(
   logic [1:0]                 HPTWRW;
   logic [2:0]                 HPTWSize;
   logic                       SelReplayMemE;
-  logic [11:0]                PreLSUAdrE;  
   logic                       ITLBMissOrDAFaultF, ITLBMissOrDAFaultNoTrapF;
   logic                       DTLBMissOrDAFaultM, DTLBMissOrDAFaultNoTrapM;
   logic                       SelHPTWAdr;
@@ -97,8 +95,10 @@ module lsuvirtmem(
 
   // Once the walk is done and it is time to update the DTLB we need to switch back 
   // to the orignal data virtual address.
-  assign SelHPTWAdr = SelHPTW & ~DTLBWriteM;
+  assign SelHPTWAdr = SelHPTW & ~(DTLBWriteM | ITLBWriteF);
-  
+
+  assign SelReplay = SelHPTWAdr | SelReplayMemE;
+
   // multiplex the outputs to LSU
   if(`XLEN+2-`PA_BITS > 0) begin
     logic [(`XLEN+2-`PA_BITS)-1:0] zeros;
@@ -109,12 +109,10 @@ module lsuvirtmem(
   mux2 #(3) sizemux(Funct3M, HPTWSize, SelHPTW, LSUFunct3M);
   mux2 #(7) funct7mux(Funct7M, 7'b0, SelHPTW, LSUFunct7M);    
   mux2 #(2) atomicmux(AtomicM, 2'b00, SelHPTW, LSUAtomicM);
-  mux2 #(12) adremux(IEUAdrE[11:0], HPTWAdr[11:0], SelHPTW, PreLSUAdrE);
   mux2 #(`XLEN+2) lsupadrmux(IEUAdrExtM, HPTWAdrExt, SelHPTWAdr, PreLSUPAdrM);
   if(`HPTW_WRITES_SUPPORTED)
     mux2 #(`XLEN) lsuwritedatamux(WriteDataM, PTE, SelHPTW, IMWriteDataM);
   else assign IMWriteDataM = WriteDataM;
-  mux2 #(12) replaymux(PreLSUAdrE, IEUAdrExtM[11:0], SelReplayMemE, LSUAdrE); // replay cpu request after hptw.  *** redudant with mux in cache.
 
   // always block interrupts when using the hardware page table walker.
   assign CPUBusy = StallW & ~SelHPTW;