Merge pull request #1355 from YosysHQ/eddie/peepopt_dffmuxext

peepopt_dffmux -- bit optimisations for word level $dff + (enable/reset) $mux cells

6 files changed, 291 insertions, 14 deletions

diff --git a/frontends/aiger/aigerparse.cc b/frontends/aiger/aigerparse.cc
index a8d5abc1e..e8ee487e5 100644
--- a/frontends/aiger/aigerparse.cc
+++ b/frontends/aiger/aigerparse.cc
@@ -985,7 +985,7 @@ void AigerReader::post_process()
 	// operate (and run checks on) this one module
 	RTLIL::Design *mapped_design = new RTLIL::Design;
 	mapped_design->add(module);
-	Pass::call(mapped_design, "clean -purge");
+	Pass::call(mapped_design, "clean");
 	mapped_design->modules_.erase(module->name);
 	delete mapped_design;
 
diff --git a/passes/pmgen/Makefile.inc b/passes/pmgen/Makefile.inc
index 4989c582a..98691d0fe 100644
--- a/passes/pmgen/Makefile.inc
+++ b/passes/pmgen/Makefile.inc
@@ -27,6 +27,7 @@ $(eval $(call add_extra_objs,passes/pmgen/peepopt_pm.h))
 
 PEEPOPT_PATTERN  = passes/pmgen/peepopt_shiftmul.pmg
 PEEPOPT_PATTERN += passes/pmgen/peepopt_muldiv.pmg
+PEEPOPT_PATTERN += passes/pmgen/peepopt_dffmux.pmg
 
 passes/pmgen/peepopt_pm.h: passes/pmgen/pmgen.py $(PEEPOPT_PATTERN)
 	$(P) mkdir -p passes/pmgen && python3 $< -o $@ -p peepopt $(filter-out $<,$^)
diff --git a/passes/pmgen/peepopt.cc b/passes/pmgen/peepopt.cc
index e7f95cf85..72b02127a 100644
--- a/passes/pmgen/peepopt.cc
+++ b/passes/pmgen/peepopt.cc
@@ -60,6 +60,7 @@ struct PeepoptPass : public Pass {
 				peepopt_pm pm(module, module->selected_cells());
 				pm.run_shiftmul();
 				pm.run_muldiv();
+				pm.run_dffmux();
 			}
 		}
 	}
diff --git a/passes/pmgen/peepopt_dffmux.pmg b/passes/pmgen/peepopt_dffmux.pmg
new file mode 100644
index 000000000..c88a52226
--- /dev/null
+++ b/passes/pmgen/peepopt_dffmux.pmg
@@ -0,0 +1,113 @@
+pattern dffmux
+
+state <IdString> cemuxAB rstmuxBA
+state <SigSpec> sigD
+
+match dff
+	select dff->type == $dff
+	select GetSize(port(dff, \D)) > 1
+endmatch
+
+match rstmux
+	select rstmux->type == $mux
+	select GetSize(port(rstmux, \Y)) > 1
+	index <SigSpec> port(rstmux, \Y) === port(dff, \D)
+	choice <IdString> BA {\B, \A}
+	select port(rstmux, BA).is_fully_const()
+	set rstmuxBA BA
+	optional
+endmatch
+
+code sigD
+	if (rstmux)
+		sigD = port(rstmux, rstmuxBA == \B ? \A : \B);
+	else
+		sigD = port(dff, \D);
+endcode
+
+match cemux
+	select cemux->type == $mux
+	select GetSize(port(cemux, \Y)) > 1
+	index <SigSpec> port(cemux, \Y) === sigD
+	choice <IdString> AB {\A, \B}
+	index <SigSpec> port(cemux, AB) === port(dff, \Q)
+	set cemuxAB AB
+endmatch
+
+code
+	SigSpec D = port(cemux, cemuxAB == \A ? \B : \A);
+	SigSpec Q = port(dff, \Q);
+	Const rst;
+	if (rstmux)
+		rst = port(rstmux, rstmuxBA).as_const();
+	int width = GetSize(D);
+
+	SigSpec &ceA = cemux->connections_.at(\A);
+	SigSpec &ceB = cemux->connections_.at(\B);
+	SigSpec &ceY = cemux->connections_.at(\Y);
+	SigSpec &dffD = dff->connections_.at(\D);
+	SigSpec &dffQ = dff->connections_.at(\Q);
+
+	if (D[width-1] == D[width-2]) {
+		did_something = true;
+
+		SigBit sign = D[width-1];
+		bool is_signed = sign.wire;
+		int i;
+		for (i = width-1; i >= 2; i--) {
+			if (!is_signed) {
+				module->connect(Q[i], sign);
+				if (D[i-1] != sign || (rst.size() && rst[i-1] != rst[width-1]))
+					break;
+			}
+			else {
+				module->connect(Q[i], Q[i-1]);
+				if (D[i-2] != sign || (rst.size() && rst[i-1] != rst[width-1]))
+					break;
+			}
+		}
+
+		ceA.remove(i, width-i);
+		ceB.remove(i, width-i);
+		ceY.remove(i, width-i);
+		cemux->fixup_parameters();
+		dffD.remove(i, width-i);
+		dffQ.remove(i, width-i);
+		dff->fixup_parameters();
+
+		log("dffcemux pattern in %s: dff=%s, cemux=%s; removed top %d bits.\n", log_id(module), log_id(dff), log_id(cemux), width-i);
+		accept;
+	}
+	else {
+		int count = 0;
+		for (int i = width-1; i >= 0; i--) {
+			if (D[i].wire)
+				continue;
+			Wire *w = Q[i].wire;
+			auto it = w->attributes.find(\init);
+			State init;
+			if (it != w->attributes.end())
+				init = it->second[Q[i].offset];
+			else
+				init = State::Sx;
+
+			if (init == State::Sx || init == D[i].data) {
+				count++;
+				module->connect(Q[i], D[i]);
+				ceA.remove(i);
+				ceB.remove(i);
+				ceY.remove(i);
+				dffD.remove(i);
+				dffQ.remove(i);
+			}
+		}
+		if (count > 0) {
+			did_something = true;
+			cemux->fixup_parameters();
+			dff->fixup_parameters();
+			log("dffcemux pattern in %s: dff=%s, cemux=%s; removed %d constant bits.\n", log_id(module), log_id(dff), log_id(cemux), count);
+		}
+
+		accept;
+	}
+endcode
diff --git a/tests/simple/peepopt.v b/tests/simple/peepopt.v
deleted file mode 100644
index 1bf427897..000000000
--- a/tests/simple/peepopt.v
+++ /dev/null
@@ -1,13 +0,0 @@
-module peepopt_shiftmul_0 #(parameter N=3, parameter W=3) (input [N*W-1:0] i, input [$clog2(N)-1:0] s, output [W-1:0] o);
-assign o = i[s*W+:W];
-endmodule
-
-module peepopt_shiftmul_1 (output y, input [2:0] w);
-assign y = 1'b1 >> (w * (3'b110));
-endmodule
-
-module peepopt_muldiv_0(input [1:0] i, output [1:0] o);
-wire [3:0] t;
-assign t = i * 3;
-assign o = t / 3;
-endmodule
diff --git a/tests/various/peepopt.ys b/tests/various/peepopt.ys
new file mode 100644
index 000000000..6bca62e2b
--- /dev/null
+++ b/tests/various/peepopt.ys
@@ -0,0 +1,175 @@
+read_verilog <<EOT
+module peepopt_shiftmul_0 #(parameter N=3, parameter W=3) (input [N*W-1:0] i, input [$clog2(N)-1:0] s, output [W-1:0] o);
+assign o = i[s*W+:W];
+endmodule
+EOT
+
+prep -nokeepdc
+equiv_opt -assert peepopt
+design -load postopt
+clean
+select -assert-count 1 t:$shiftx
+select -assert-count 0 t:$shiftx t:* %D
+
+####################
+
+design -reset
+read_verilog <<EOT
+module peepopt_shiftmul_1 (output [7:0] y, input [2:0] w);
+assign y = 1'b1 >> (w * (3'b110));
+endmodule
+EOT
+
+prep -nokeepdc
+equiv_opt -assert peepopt
+design -load postopt
+clean
+select -assert-count 1 t:$shr
+select -assert-count 1 t:$mul
+select -assert-count 0 t:$shr t:$mul %% t:* %D
+
+####################
+
+design -reset
+read_verilog <<EOT
+module peepopt_shiftmul_2 (input [11:0] D, input [1:0] S, output [11:0] Y);
+	assign Y = D >> (S*3);
+endmodule
+EOT
+
+prep
+design -save gold
+peepopt
+design -stash gate
+
+design -import gold -as gold peepopt_shiftmul_2
+design -import gate -as gate peepopt_shiftmul_2
+
+miter -equiv -make_assert -make_outputs -ignore_gold_x -flatten gold gate miter
+sat -show-public -enable_undef -prove-asserts miter
+cd gate
+select -assert-count 1 t:$shr
+select -assert-count 1 t:$mul
+select -assert-count 0 t:$shr t:$mul %% t:* %D
+
+####################
+
+design -reset
+read_verilog <<EOT
+module peepopt_muldiv_0(input [1:0] i, output [1:0] o);
+wire [3:0] t;
+assign t = i * 3;
+assign o = t / 3;
+endmodule
+EOT
+
+prep -nokeepdc
+equiv_opt -assert peepopt
+design -load postopt
+clean
+select -assert-count 0 t:*
+
+####################
+
+design -reset
+read_verilog <<EOT
+module peepopt_dffmuxext_unsigned(input clk, ce, input [1:0] i, output reg [3:0] o);
+    always @(posedge clk) if (ce) o <= i;
+endmodule
+EOT
+
+proc
+equiv_opt -assert peepopt
+design -load postopt
+clean
+select -assert-count 1 t:$dff r:WIDTH=2 %i
+select -assert-count 1 t:$mux r:WIDTH=2 %i
+select -assert-count 0 t:$dff t:$mux %% t:* %D
+
+####################
+
+design -reset
+read_verilog <<EOT
+module peepopt_dffmuxext_signed(input clk, ce, input signed [1:0] i, output reg signed [3:0] o);
+    always @(posedge clk) if (ce) o <= i;
+endmodule
+EOT
+
+proc
+equiv_opt -assert peepopt
+design -load postopt
+clean
+select -assert-count 1 t:$dff r:WIDTH=2 %i
+select -assert-count 1 t:$mux r:WIDTH=2 %i
+select -assert-count 0 t:$dff t:$mux %% t:* %D
+
+###################
+
+design -reset
+read_verilog <<EOT
+module peepopt_dffmuxext_const(input clk, ce, input [1:0] i, output reg [5:0] o);
+    always @(posedge clk) if (ce) o <= {1'b0, i[1], 2'b1x, i[0], 1'bz};
+endmodule
+EOT
+
+proc
+equiv_opt -assert peepopt
+design -load postopt
+select -assert-count 1 t:$dff r:WIDTH=2 %i
+select -assert-count 1 t:$mux r:WIDTH=2 %i
+select -assert-count 0 t:$dff t:$mux %% t:* %D
+
+###################
+
+design -reset
+read_verilog <<EOT
+module peepopt_dffmuxext_const_init(input clk, ce, input [1:0] i, (* init=6'b0x00x1 *) output reg [5:0] o);
+    always @(posedge clk) if (ce) o <= {1'b0, i[1], 2'b1x, i[0], 1'bz};
+endmodule
+EOT
+
+proc
+equiv_opt -assert peepopt
+design -load postopt
+select -assert-count 1 t:$dff r:WIDTH=5 %i
+select -assert-count 1 t:$mux r:WIDTH=5 %i
+select -assert-count 0 t:$dff t:$mux %% t:* %D
+
+####################
+
+design -reset
+read_verilog <<EOT
+module peepopt_dffmuxext_unsigned_rst(input clk, ce, rst, input [1:0] i, output reg [3:0] o);
+    always @(posedge clk) if (rst) o <= 0; else if (ce) o <= i;
+endmodule
+EOT
+
+proc
+equiv_opt -assert peepopt
+design -load postopt
+wreduce
+select -assert-count 1 t:$dff r:WIDTH=2 %i
+select -assert-count 2 t:$mux
+select -assert-count 2 t:$mux r:WIDTH=2 %i
+select -assert-count 0 t:$dff t:$mux %% t:* %D
+
+####################
+
+design -reset
+read_verilog <<EOT
+module peepopt_dffmuxext_signed_rst(input clk, ce, rstn, input signed [1:0] i, output reg signed [3:0] o);
+    always @(posedge clk) begin
+        if (ce) o <= i;
+        if (!rstn) o <= 4'b1111;
+    end
+endmodule
+EOT
+
+proc
+equiv_opt -assert peepopt
+design -load postopt
+wreduce
+select -assert-count 1 t:$dff r:WIDTH=2 %i
+select -assert-count 2 t:$mux
+select -assert-count 2 t:$mux r:WIDTH=2 %i
+select -assert-count 0 t:$logic_not t:$dff t:$mux %% t:* %D