From 5447159cfd03801dc4f520885b719c5c4c3a91e4 Mon Sep 17 00:00:00 2001
From: Rose Thompson <ross1728@gmail.com>
Date: Wed, 6 Mar 2024 15:15:26 -0600
Subject: [PATCH] Revert "Cleanup."

This reverts commit e84b7cc14782d6cb6676d94e987c15f63f87a604.
---
 src/lsu/align.sv      | 18 ++++++++++++++++--
 src/lsu/endianswap.sv |  4 ++--
 src/lsu/lsu.sv        |  4 +++-
 3 files changed, 21 insertions(+), 5 deletions(-)

diff --git a/src/lsu/align.sv b/src/lsu/align.sv
index 094da4a15..fa10916f9 100644
--- a/src/lsu/align.sv
+++ b/src/lsu/align.sv
@@ -48,6 +48,7 @@ module align import cvw::*;  #(parameter cvw_t P) (
   input logic [P.LLEN-1:0]        LSUWriteDataM, 
 
   output logic [(P.LLEN*2-1)/8:0] ByteMaskSpillM,
+  output logic [P.LLEN*2-1:0]     LSUWriteDataSpillM, 
 
   output logic [P.XLEN-1:0]       IEUAdrSpillE, // The next PCF for one of the two memory addresses of the spill
   output logic [P.XLEN-1:0]       IEUAdrSpillM, // IEUAdrM for one of the two memory addresses of the spill
@@ -70,8 +71,9 @@ module align import cvw::*;  #(parameter cvw_t P) (
   logic [P.XLEN-1:0]   IEUAdrIncrementM;
 
   localparam OFFSET_LEN = $clog2(LLENINBYTES);
-  logic [OFFSET_LEN-1:0] AccessByteOffsetM;
-  logic                  PotentialSpillM;
+  logic [$clog2(LLENINBYTES)-1:0]              AccessByteOffsetM;
+  logic [$clog2(LLENINBYTES)+2:0]              ShiftAmount;
+  logic                                        PotentialSpillM;
 
   /* verilator lint_off WIDTHEXPAND */
   assign IEUAdrIncrementM = IEUAdrM + LLENINBYTES;
@@ -140,6 +142,18 @@ module align import cvw::*;  #(parameter cvw_t P) (
   // merge together
   mux2 #(2*P.LLEN) postspillmux(DCacheReadDataWordM, {DCacheReadDataWordM[P.LLEN-1:0], ReadDataWordFirstHalfM}, SelSpillM, ReadDataWordSpillAllM);
 
+
+  // shifter (4:1 mux for 32 bit, 8:1 mux for 64 bit)
+  // 8 * is for shifting by bytes not bits
+  assign ShiftAmount = SelHPTW ? '0 : {AccessByteOffsetM, 3'b0}; // AND gate
+
+  // write path. Also has the 8:1 shifter muxing for the byteoffset
+  // then it also has the mux to select when a spill occurs
+  logic [P.LLEN*3-1:0] LSUWriteDataShiftedExtM;  // *** RT: Find a better way.  I've extending in both directions so we don't shift in zeros.  The cache expects the writedata to not have any zero data, but instead replicated data.
+
+  assign LSUWriteDataShiftedExtM = {LSUWriteDataM, LSUWriteDataM, LSUWriteDataM} << ShiftAmount;
+  assign LSUWriteDataSpillM = LSUWriteDataShiftedExtM[P.LLEN*3-1:P.LLEN];
+
   mux3 #(2*P.LLEN/8) bytemaskspillmux({ByteMaskExtendedM, ByteMaskM}, // no spill
                                       {{{P.LLEN/8}{1'b0}}, ByteMaskM}, // spill, first half
                                       {{{P.LLEN/8}{1'b0}}, ByteMaskExtendedM}, // spill, second half
diff --git a/src/lsu/endianswap.sv b/src/lsu/endianswap.sv
index afd4ecdd2..3c552b371 100644
--- a/src/lsu/endianswap.sv
+++ b/src/lsu/endianswap.sv
@@ -54,8 +54,8 @@ module endianswap #(parameter LEN) (
         y[143:136] = a[119:112];
         y[135:128] = a[127:120];
         y[127:120] = a[135:128];
-        y[119:112] = a[143:136];
-        y[111:104] = a[151:144];
+        y[119:112] = a[142:136];
+        y[111:104] = a[152:144];
         y[103:96]  = a[159:152];
         y[95:88]   = a[167:160];
         y[87:80]   = a[175:168];
diff --git a/src/lsu/lsu.sv b/src/lsu/lsu.sv
index cf1767f5c..e3e138ea4 100644
--- a/src/lsu/lsu.sv
+++ b/src/lsu/lsu.sv
@@ -126,6 +126,7 @@ module lsu import cvw::*;  #(parameter cvw_t P) (
   logic [P.LLEN-1:0]     DTIMReadDataWordM;                      // DTIM read data
   /* verilator lint_off WIDTHEXPAND */  
   logic [MLEN-1:0]       DCacheReadDataWordM;                    // D$ read data
+  logic [MLEN-1:0]       LSUWriteDataSpillM;                     // Final write data
   logic [MLEN/8-1:0]     ByteMaskSpillM;                         // Selects which bytes within a word to write
   /* verilator lint_on WIDTHEXPAND */
   logic [MLEN-1:0]       ReadDataWordMuxM;                       // DTIM or D$ read data
@@ -166,7 +167,7 @@ module lsu import cvw::*;  #(parameter cvw_t P) (
     align #(P) align(.clk, .reset, .StallM, .FlushM, .IEUAdrE, .IEUAdrM, .Funct3M, .FpLoadStoreM, 
                      .MemRWM,
                      .DCacheReadDataWordM, .CacheBusHPWTStall, .SelHPTW,
-                     .ByteMaskM, .ByteMaskExtendedM, .LSUWriteDataM(LSUWriteDataM[P.LLEN-1:0]), .ByteMaskSpillM,
+                     .ByteMaskM, .ByteMaskExtendedM, .LSUWriteDataM(LSUWriteDataM[P.LLEN-1:0]), .ByteMaskSpillM, .LSUWriteDataSpillM,
                      .IEUAdrSpillE, .IEUAdrSpillM, .SelSpillE, .ReadDataWordSpillAllM, .SpillStallM);
     assign IEUAdrExtM = {2'b00, IEUAdrSpillM}; 
     assign IEUAdrExtE = {2'b00, IEUAdrSpillE};
@@ -176,6 +177,7 @@ module lsu import cvw::*;  #(parameter cvw_t P) (
     assign SelSpillE = '0;
     assign ReadDataWordSpillAllM = DCacheReadDataWordM;
     assign ByteMaskSpillM = ByteMaskM;
+    assign LSUWriteDataSpillM = LSUWriteDataM;
     assign MemRWSpillM = MemRWM;
     assign {SpillStallM} = '0;
   end