fixed snoop forwarding bug and single bank support

hw/rtl/cache/VX_snp_forwarder.v

@@ -34,11 +34,13 @@ module VX_snp_forwarder #(
     output wire [NUM_REQUESTS-1:0]      snp_fwdin_ready
 );
     reg [`DRAM_ADDR_WIDTH+SNP_REQ_TAG_WIDTH-1:0] pending_reqs [SNRQ_SIZE-1:0];
-    reg [`REQS_BITS-1:0] pending_cntrs [SNRQ_SIZE-1:0];
-    reg [`LOG2UP(SNRQ_SIZE)-1:0] rd_ptr, wr_ptr;
-    reg [`LOG2UP(SNRQ_SIZE)-1:0] pending_size;
+    reg [`REQS_BITS:0] pending_cntrs [SNRQ_SIZE-1:0];
+    reg [`LOG2UP(SNRQ_SIZE):0] rd_ptr, wr_ptr;
     reg [`REQS_BITS-1:0] fwdin_sel;
-    wire enqueue, dequeue, empty;
+
+    wire [`LOG2UP(SNRQ_SIZE)-1:0] rd_a, wr_a;
+
+    wire enqueue, dequeue, empty, full;
 
     wire fwdout_ready;
 
@@ -49,47 +51,43 @@ module VX_snp_forwarder #(
     
     assign fwdout_ready = (& snp_fwdout_ready);
 
-    assign snp_req_ready = (pending_size != `LOG2UP(SNRQ_SIZE)'(SNRQ_SIZE-1)) // not full
-                        && fwdout_ready;
+    assign snp_req_ready = !full && fwdout_ready;
+
+    assign rd_a = rd_ptr[`LOG2UP(SNRQ_SIZE)-1:0];
+    assign wr_a = wr_ptr[`LOG2UP(SNRQ_SIZE)-1:0];
 
     genvar i;
 
     for (i = 0; i < NUM_REQUESTS; i++) begin
         assign snp_fwdout_valid[i] = enqueue && fwdout_ready;
         assign snp_fwdout_addr[i]  = snp_req_addr;
-        assign snp_fwdout_tag[i]   = wr_ptr;
+        assign snp_fwdout_tag[i]   = wr_a;
     end
 
     assign fwdin_ready = snp_rsp_ready;
-
     assign fwdin_taken = fwdin_valid && fwdin_ready;
 
     assign snp_rsp_valid = fwdin_taken && (1 == pending_cntrs[fwdin_tag]); // send response
     assign {snp_rsp_addr, snp_rsp_tag} = pending_reqs[fwdin_tag];
 
     assign empty   = (wr_ptr == rd_ptr);
+    assign full    = (wr_a == rd_a) && (wr_ptr[`LOG2UP(SNRQ_SIZE)] != rd_ptr[`LOG2UP(SNRQ_SIZE)]);
+    
     assign enqueue = snp_req_valid && snp_req_ready;       
-    assign dequeue = !empty && (0 == pending_cntrs[rd_ptr]);
+    assign dequeue = !empty && (0 == pending_cntrs[rd_a]);
 
     always @(posedge clk) begin
         if (reset) begin
-            rd_ptr       <= 0;
-            wr_ptr       <= 0;
-            pending_size <= 0;
+            rd_ptr <= 0;
+            wr_ptr <= 0;
         end else begin
             if (enqueue)  begin
-                pending_reqs[wr_ptr]  <= {snp_req_addr, snp_req_tag};
-                pending_cntrs[wr_ptr] <= `REQS_BITS'(NUM_REQUESTS);
-                wr_ptr                <= wr_ptr + 1;
-                if (!dequeue)  begin
-                    pending_size <= pending_size + 1;
-                end
+                pending_reqs[wr_a]  <= {snp_req_addr, snp_req_tag};
+                pending_cntrs[wr_a] <= NUM_REQUESTS;
+                wr_ptr              <= wr_ptr + 1;
             end            
             if (dequeue) begin
                 rd_ptr <= rd_ptr + 1;
-                if (!enqueue)  begin
-                    pending_size <= pending_size - 1;
-                end
             end
             if (fwdin_taken) begin
                 pending_cntrs[fwdin_tag] <= pending_cntrs[fwdin_tag] - 1;