From fc158689adb706979c42f8ebff9a7369a717d25b Mon Sep 17 00:00:00 2001
From: David Harris <David_Harris@hmc.edu>
Date: Sun, 24 Mar 2024 17:15:46 -0700
Subject: [PATCH] Shared amoalu max/min comparator hardware and removed input
 sign extend muxes

---
 src/lsu/amoalu.sv | 31 +++++++++++++++----------------
 1 file changed, 15 insertions(+), 16 deletions(-)

diff --git a/src/lsu/amoalu.sv b/src/lsu/amoalu.sv
index f4b12d4af..a6ee53e73 100644
--- a/src/lsu/amoalu.sv
+++ b/src/lsu/amoalu.sv
@@ -37,12 +37,16 @@ module amoalu import cvw::*;  #(parameter cvw_t P) (
 );
 
   logic [P.XLEN-1:0] a, b, y;
-  logic              sngd, eq, lt, eq32, lt32, w64;
+  logic               lt, cmp, sngd, sngd32, eq32, lt32, w64;
+
+  // Rename inputs
+  assign a = ReadDataM;
+  assign b = IHWriteDataM;
 
   // Share hardware among the four amomin/amomax comparators
   assign sngd = ~LSUFunct7M[5]; // Funct7[5] = 0 for signed amomin/max
   assign w64 = (LSUFunct3M[1:0] == 2'b10); // operate on bottom 32 bits
-  assign sngd32 = sngd & (P.XLEN == 32 | w64);
+  assign sngd32 = sngd & (P.XLEN == 32 | w64); // flip sign in lower 32 bits on 32-bit comparisons only
 
   comparator #(32) cmp32(a[31:0], b[31:0], sngd32, {eq32, lt32});
   if (P.XLEN == 32) begin
@@ -51,12 +55,13 @@ module amoalu import cvw::*;  #(parameter cvw_t P) (
     logic equpper, ltupper, lt64;
 
     comparator #(32) cmpupper(a[63:32], b[63:32], sngd, {equpper, ltupper});
-    
     assign lt64 = ltupper | equpper & lt32;
     assign lt = w64 ? lt32 : lt64;
   end
 
-  // and the same mux can be used to select b for swap.
+  assign cmp = lt ^ LSUFunct7M[4]; // flip sense of comparison for maximums
+
+  // AMO ALU
   always_comb 
     case (LSUFunct7M[6:2])
       5'b00001: y = b;           // amoswap
@@ -64,27 +69,21 @@ module amoalu import cvw::*;  #(parameter cvw_t P) (
       5'b00100: y = a ^ b;       // amoxor
       5'b01100: y = a & b;       // amoand
       5'b01000: y = a | b;       // amoor
-      5'b10000: y = lt ? a : b;  // amomin
-      5'b10100: y = lt ? b : a;  // amomax
-      5'b11000: y = lt ? a : b;  // amominu
-      5'b11100: y = lt ? b : a;  // amomaxu
+      5'b10000: y = cmp ? a : b; // amomin
+      5'b10100: y = cmp ? a : b; // amomax
+      5'b11000: y = cmp ? a : b; // amominu
+      5'b11100: y = cmp ? a : b; // amomaxu
       default:  y = 'x;          // undefined; *** could change to b for efficiency
     endcase
 
-  // sign extend if necessary
+  // sign extend output if necessary for w64
   if (P.XLEN == 32) begin:sext
-    assign a = ReadDataM;
-    assign b = IHWriteDataM;
     assign AMOResultM = y;
   end else begin:sext // P.XLEN = 64
     always_comb 
-      if (LSUFunct3M[1:0] == 2'b10) begin // sign-extend word-length operations
-        a = {{32{ReadDataM[31]}}, ReadDataM[31:0]};
-        b = {{32{IHWriteDataM[31]}}, IHWriteDataM[31:0]};
+      if (w64) begin // sign-extend word-length operations
         AMOResultM = {{32{y[31]}}, y[31:0]};
       end else begin
-        a = ReadDataM;
-        b = IHWriteDataM;
         AMOResultM = y;
       end
   end