rtl/perf: add a flag for enabling or disabling the perf module

3 files changed, 223 insertions, 208 deletions

diff --git a/rtl/perf/perf_link.sv b/rtl/perf/perf_link.sv
index 323af45..d42fc50 100644
--- a/rtl/perf/perf_link.sv
+++ b/rtl/perf/perf_link.sv
@@ -1,4 +1,5 @@
 `include "cache/defs.sv"
+`include "config.sv"
 
 module perf_link
 (
@@ -55,127 +56,131 @@ module perf_link
 		.*
 	);
 
-	assign cached = snoop_addr_bits.io == `IO_CACHED;
-	assign mem_cycles = mem_cycles_hold + 1;
-	assign snoop_addr_bits = snoop_address;
-
-	always_comb
-		if (!address[3]) unique case (address[2:0])
-			3'b000: readdata = reads;
-			3'b001: readdata = writes;
-			3'b010: readdata = {max_read_cycles, min_read_cycles};
-			3'b011: readdata = {max_write_cycles, min_write_cycles};
-			3'b100: readdata = ring_reads;
-			3'b101: readdata = ring_invals;
-			3'b110: readdata = ring_read_invals;
-			3'b111: readdata = ring_replies;
-		endcase else unique case (address[1:0])
-			2'b00:  readdata = ring_forwards;
-			2'b01:  readdata = {max_ring_cycles, min_ring_cycles};
-			2'b10:  readdata = io_reads;
-			2'b11:  readdata = io_writes;
-		endcase
-
-	always @(posedge clk or negedge rst_n)
-		if (!rst_n) begin
-			reads <= 0;
-			writes <= 0;
-			io_reads <= 0;
-			io_writes <= 0;
-
-			min_ring_cycles <= 0;
-			max_ring_cycles <= 0;
-			min_read_cycles <= 0;
-			max_read_cycles <= 0;
-			min_write_cycles <= 0;
-			max_write_cycles <= 0;
-
-			ring_reads <= 0;
-			ring_invals <= 0;
-			ring_replies <= 0;
-			ring_forwards <= 0;
-			ring_read_invals <= 0;
-
-			mem_cycles_hold <= 0;
-		end else begin
-			ring_cycles <= ring_cycles + 1;
-
-			if (mem_read || mem_write)
-				mem_cycles_hold <= mem_cycles;
-
-			if ((mem_read || mem_write) && !mem_waitrequest) begin
-				mem_cycles_hold <= 0;
-
-				if (!cached) begin
-					if (mem_write)
-						io_writes <= io_writes + 1;
-					else
-						io_reads <= io_reads + 1;
-				end else if (mem_write) begin
-					writes <= writes + 1;
-
-					if (min_write_cycles == 0 || mem_cycles_hold < min_write_cycles)
-						min_write_cycles <= mem_cycles;
-
-					if (mem_cycles_hold >= max_write_cycles)
-						max_write_cycles <= mem_cycles;
+	generate
+		if (`CONFIG_PERF_MONITOR) begin: enable
+			assign cached = snoop_addr_bits.io == `IO_CACHED;
+			assign mem_cycles = mem_cycles_hold + 1;
+			assign snoop_addr_bits = snoop_address;
+
+			always_comb
+				if (!address[3]) unique case (address[2:0])
+					3'b000: readdata = reads;
+					3'b001: readdata = writes;
+					3'b010: readdata = {max_read_cycles, min_read_cycles};
+					3'b011: readdata = {max_write_cycles, min_write_cycles};
+					3'b100: readdata = ring_reads;
+					3'b101: readdata = ring_invals;
+					3'b110: readdata = ring_read_invals;
+					3'b111: readdata = ring_replies;
+				endcase else unique case (address[1:0])
+					2'b00:  readdata = ring_forwards;
+					2'b01:  readdata = {max_ring_cycles, min_ring_cycles};
+					2'b10:  readdata = io_reads;
+					2'b11:  readdata = io_writes;
+				endcase
+
+			always @(posedge clk or negedge rst_n)
+				if (!rst_n) begin
+					reads <= 0;
+					writes <= 0;
+					io_reads <= 0;
+					io_writes <= 0;
+
+					min_ring_cycles <= 0;
+					max_ring_cycles <= 0;
+					min_read_cycles <= 0;
+					max_read_cycles <= 0;
+					min_write_cycles <= 0;
+					max_write_cycles <= 0;
+
+					ring_reads <= 0;
+					ring_invals <= 0;
+					ring_replies <= 0;
+					ring_forwards <= 0;
+					ring_read_invals <= 0;
+
+					mem_cycles_hold <= 0;
 				end else begin
-					reads <= reads + 1;
-
-					if (min_read_cycles == 0 || mem_cycles_hold < min_read_cycles)
-						min_read_cycles <= mem_cycles;
-
-					if (mem_cycles_hold >= max_read_cycles)
-						max_read_cycles <= mem_cycles;
+					ring_cycles <= ring_cycles + 1;
+
+					if (mem_read || mem_write)
+						mem_cycles_hold <= mem_cycles;
+
+					if ((mem_read || mem_write) && !mem_waitrequest) begin
+						mem_cycles_hold <= 0;
+
+						if (!cached) begin
+							if (mem_write)
+								io_writes <= io_writes + 1;
+							else
+								io_reads <= io_reads + 1;
+						end else if (mem_write) begin
+							writes <= writes + 1;
+
+							if (min_write_cycles == 0 || mem_cycles_hold < min_write_cycles)
+								min_write_cycles <= mem_cycles;
+
+							if (mem_cycles_hold >= max_write_cycles)
+								max_write_cycles <= mem_cycles;
+						end else begin
+							reads <= reads + 1;
+
+							if (min_read_cycles == 0 || mem_cycles_hold < min_read_cycles)
+								min_read_cycles <= mem_cycles;
+
+							if (mem_cycles_hold >= max_read_cycles)
+								max_read_cycles <= mem_cycles;
+						end
+					end
+
+					if (snoop_left_valid && snoop_left_ready && snoop_left.ttl == `TTL_END) begin
+						if (snoop_left.reply)
+							ring_replies <= ring_replies + 1;
+
+						if (min_ring_cycles == 0 || ring_cycles < min_ring_cycles)
+							min_ring_cycles <= ring_cycles;
+
+						if (ring_cycles > max_ring_cycles)
+							max_ring_cycles <= ring_cycles;
+					end
+
+					if (snoop_right_valid && snoop_right_ready) begin
+						if (snoop_right.ttl == `TTL_MAX) begin
+							ring_cycles <= 1;
+
+							if (snoop_right.read && !snoop_right.inval)
+								ring_reads <= ring_reads + 1;
+
+							if (!snoop_right.read && snoop_right.inval)
+								ring_invals <= ring_invals + 1;
+
+							if (snoop_right.read && snoop_right.inval)
+								ring_read_invals <= ring_read_invals + 1;
+						end else
+							ring_forwards <= ring_forwards + 1;
+					end
+
+					if (clear) begin
+						reads <= 0;
+						writes <= 0;
+						io_reads <= 0;
+						io_writes <= 0;
+
+						min_ring_cycles <= 0;
+						max_ring_cycles <= 0;
+						min_read_cycles <= 0;
+						max_read_cycles <= 0;
+						min_write_cycles <= 0;
+						max_write_cycles <= 0;
+
+						ring_reads <= 0;
+						ring_invals <= 0;
+						ring_replies <= 0;
+						ring_forwards <= 0;
+						ring_read_invals <= 0;
+					end
 				end
-			end
-
-			if (snoop_left_valid && snoop_left_ready && snoop_left.ttl == `TTL_END) begin
-				if (snoop_left.reply)
-					ring_replies <= ring_replies + 1;
-
-				if (min_ring_cycles == 0 || ring_cycles < min_ring_cycles)
-					min_ring_cycles <= ring_cycles;
-
-				if (ring_cycles > max_ring_cycles)
-					max_ring_cycles <= ring_cycles;
-			end
-
-			if (snoop_right_valid && snoop_right_ready) begin
-				if (snoop_right.ttl == `TTL_MAX) begin
-					ring_cycles <= 1;
-
-					if (snoop_right.read && !snoop_right.inval)
-						ring_reads <= ring_reads + 1;
-
-					if (!snoop_right.read && snoop_right.inval)
-						ring_invals <= ring_invals + 1;
-
-					if (snoop_right.read && snoop_right.inval)
-						ring_read_invals <= ring_read_invals + 1;
-				end else
-					ring_forwards <= ring_forwards + 1;
-			end
-
-			if (clear) begin
-				reads <= 0;
-				writes <= 0;
-				io_reads <= 0;
-				io_writes <= 0;
-
-				min_ring_cycles <= 0;
-				max_ring_cycles <= 0;
-				min_read_cycles <= 0;
-				max_read_cycles <= 0;
-				min_write_cycles <= 0;
-				max_write_cycles <= 0;
-
-				ring_reads <= 0;
-				ring_invals <= 0;
-				ring_replies <= 0;
-				ring_forwards <= 0;
-				ring_read_invals <= 0;
-			end
 		end
+	endgenerate
 
 endmodule
diff --git a/rtl/perf/perf_monitor.sv b/rtl/perf/perf_monitor.sv
index 2f38d94..b7bc6ba 100644
--- a/rtl/perf/perf_monitor.sv
+++ b/rtl/perf/perf_monitor.sv
@@ -1,4 +1,5 @@
 `include "cache/defs.sv"
+`include "config.sv"
 
 module perf_monitor
 (
@@ -252,35 +253,39 @@ module perf_monitor
 		.*
 	);
 
-	assign address = perf_address[3:0];
-
-	always_comb begin
-		clear_0 = 0;
-		clear_1 = 0;
-		clear_2 = 0;
-		clear_3 = 0;
-
-		unique case (perf_address[5:4])
-			2'b00: begin
-				clear_0 = perf_write;
-				perf_readdata = readdata_0;
-			end
-
-			2'b01: begin
-				clear_1 = perf_write;
-				perf_readdata = readdata_1;
-			end
-
-			2'b10: begin
-				clear_2 = perf_write;
-				perf_readdata = readdata_2;
-			end
-
-			2'b11: begin
-				clear_3 = perf_write;
-				perf_readdata = readdata_3;
+	generate
+		if (`CONFIG_PERF_MONITOR) begin: enable
+			assign address = perf_address[3:0];
+
+			always_comb begin
+				clear_0 = 0;
+				clear_1 = 0;
+				clear_2 = 0;
+				clear_3 = 0;
+
+				unique case (perf_address[5:4])
+					2'b00: begin
+						clear_0 = perf_write;
+						perf_readdata = readdata_0;
+					end
+
+					2'b01: begin
+						clear_1 = perf_write;
+						perf_readdata = readdata_1;
+					end
+
+					2'b10: begin
+						clear_2 = perf_write;
+						perf_readdata = readdata_2;
+					end
+
+					2'b11: begin
+						clear_3 = perf_write;
+						perf_readdata = readdata_3;
+					end
+				endcase
 			end
-		endcase
-	end
+		end
+	endgenerate
 
 endmodule
diff --git a/rtl/perf/perf_snoop.sv b/rtl/perf/perf_snoop.sv
index e98153e..2cc159d 100644
--- a/rtl/perf/perf_snoop.sv
+++ b/rtl/perf/perf_snoop.sv
@@ -1,4 +1,5 @@
 `include "cache/defs.sv"
+`include "config.sv"
 
 module perf_snoop
 (
@@ -64,66 +65,70 @@ module perf_snoop
 	assign local_readdata = mem_readdata;
 	assign local_waitrequest = mem_waitrequest;
 
-	always @(posedge clk or negedge rst_n)
-		if (!rst_n) begin
-			hold_read <= 0;
-			hold_write <= 0;
-			hold_waitrequest <= 0;
-
-			hold_left_ready <= 0;
-			hold_left_valid <= 0;
-			hold_right_ready <= 0;
-			hold_right_valid <= 0;
-
-			snoop_read <= 0;
-			snoop_write <= 0;
-			snoop_waitrequest <= 0;
-
-			snoop_left_ready <= 0;
-			snoop_left_valid <= 0;
-			snoop_right_ready <= 0;
-			snoop_right_valid <= 0;
-		end else begin
-			/* La idea aquí es aligerar el trabajo del fitter, ya que perf_monitor
-			 * muestrea el anillo completo, por lo que su span de área es
-			 * potencialmente grande.
-			 */
-
-			hold_read <= mem_read;
-			hold_write <= mem_write;
-			hold_waitrequest <= mem_waitrequest;
-
-			hold_left_ready <= in_left_ready;
-			hold_left_valid <= in_left_valid;
-			hold_right_ready <= in_right_ready;
-			hold_right_valid <= in_right_valid;
-
-			snoop_read <= hold_read;
-			snoop_write <= hold_write;
-			snoop_waitrequest <= hold_waitrequest;
-
-			snoop_left_ready <= hold_left_ready;
-			snoop_left_valid <= hold_left_valid;
-			snoop_right_ready <= hold_right_ready;
-			snoop_right_valid <= hold_right_valid;
+	generate
+		if (`CONFIG_PERF_MONITOR) begin: enable
+			always @(posedge clk or negedge rst_n)
+				if (!rst_n) begin
+					hold_read <= 0;
+					hold_write <= 0;
+					hold_waitrequest <= 0;
+
+					hold_left_ready <= 0;
+					hold_left_valid <= 0;
+					hold_right_ready <= 0;
+					hold_right_valid <= 0;
+
+					snoop_read <= 0;
+					snoop_write <= 0;
+					snoop_waitrequest <= 0;
+
+					snoop_left_ready <= 0;
+					snoop_left_valid <= 0;
+					snoop_right_ready <= 0;
+					snoop_right_valid <= 0;
+				end else begin
+					/* La idea aquí es aligerar el trabajo del fitter, ya que perf_monitor
+					 * muestrea el anillo completo, por lo que su span de área es
+					 * potencialmente grande.
+					 */
+
+					hold_read <= mem_read;
+					hold_write <= mem_write;
+					hold_waitrequest <= mem_waitrequest;
+
+					hold_left_ready <= in_left_ready;
+					hold_left_valid <= in_left_valid;
+					hold_right_ready <= in_right_ready;
+					hold_right_valid <= in_right_valid;
+
+					snoop_read <= hold_read;
+					snoop_write <= hold_write;
+					snoop_waitrequest <= hold_waitrequest;
+
+					snoop_left_ready <= hold_left_ready;
+					snoop_left_valid <= hold_left_valid;
+					snoop_right_ready <= hold_right_ready;
+					snoop_right_valid <= hold_right_valid;
+				end
+
+			always @(posedge clk) begin
+				hold_left.ttl <= in_left.ttl;
+				hold_left.read <= in_left.read;
+				hold_left.inval <= in_left.inval;
+				hold_left.reply <= in_left.reply;
+
+				hold_right.ttl <= in_right.ttl;
+				hold_right.read <= in_right.read;
+				hold_right.inval <= in_right.inval;
+				hold_right.reply <= in_right.reply;
+
+				snoop_left <= hold_left;
+				snoop_right <= hold_right;
+
+				hold_address <= mem_address;
+				snoop_address <= hold_address;
+			end
 		end
-
-	always @(posedge clk) begin
-		hold_left.ttl <= in_left.ttl;
-		hold_left.read <= in_left.read;
-		hold_left.inval <= in_left.inval;
-		hold_left.reply <= in_left.reply;
-
-		hold_right.ttl <= in_right.ttl;
-		hold_right.read <= in_right.read;
-		hold_right.inval <= in_right.inval;
-		hold_right.reply <= in_right.reply;
-
-		snoop_left <= hold_left;
-		snoop_right <= hold_right;
-
-		hold_address <= mem_address;
-		snoop_address <= hold_address;
-	end
+	endgenerate
 
 endmodule