Merge branch 'master' of github.com:YosysHQ/yosys

6 files changed, 162 insertions, 22 deletions

diff --git a/techlibs/anlogic/cells_map.v b/techlibs/anlogic/cells_map.v
index f54a81dcc..cfc743a4b 100644
--- a/techlibs/anlogic/cells_map.v
+++ b/techlibs/anlogic/cells_map.v
@@ -1,19 +1,19 @@
-module  \$_DFF_N_ (input D, C, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(1'b0)); endmodule
-module  \$_DFF_P_ (input D, C, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(1'b0)); endmodule
+module  \$_DFF_N_ (input D, C, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'bx), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(1'b0)); endmodule
+module  \$_DFF_P_ (input D, C, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'bx), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(1'b0)); endmodule
 
-module  \$_DFFE_NN_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(E), .sr(1'b0)); endmodule
-module  \$_DFFE_NP_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(E), .sr(1'b0)); endmodule
-module  \$_DFFE_PN_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(E), .sr(1'b0)); endmodule
-module  \$_DFFE_PP_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(E), .sr(1'b0)); endmodule
+module  \$_DFFE_NN_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'bx), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(E), .sr(1'b0)); endmodule
+module  \$_DFFE_NP_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'bx), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(E), .sr(1'b0)); endmodule
+module  \$_DFFE_PN_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'bx), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(E), .sr(1'b0)); endmodule
+module  \$_DFFE_PP_ (input D, C, E, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'bx), .SRMUX("SR"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(E), .sr(1'b0)); endmodule
 
-module  \$_DFF_NN0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
-module  \$_DFF_NN1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("SET"),   .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
-module  \$_DFF_NP0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"),  .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
-module  \$_DFF_NP1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("SET"),   .SRMUX("SR"),  .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
-module  \$_DFF_PN0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C) , .ce(1'b1), .sr(R)); endmodule
-module  \$_DFF_PN1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("SET"),   .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(R)); endmodule
-module  \$_DFF_PP0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("RESET"), .SRMUX("SR"),  .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(R)); endmodule
-module  \$_DFF_PP1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET("SET"),   .SRMUX("SR"), . SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_NN0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b0), .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_NN1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b1), .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_NP0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b0), .SRMUX("SR"),  .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_NP1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b1), .SRMUX("SR"),  .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(~C), .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_PN0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b0), .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C) , .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_PN1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b1), .SRMUX("INV"), .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_PP0_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b0), .SRMUX("SR"),  .SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(R)); endmodule
+module  \$_DFF_PP1_ (input D, C, R, output Q); AL_MAP_SEQ #(.DFFMODE("FF"), .REGSET(1'b1), .SRMUX("SR"), . SRMODE("SYNC")) _TECHMAP_REPLACE_ (.d(D), .q(Q), .clk(C),  .ce(1'b1), .sr(R)); endmodule
 
 module \$_DLATCH_N_ (E, D, Q);
   wire [1023:0] _TECHMAP_DO_ = "simplemap; opt";
diff --git a/techlibs/anlogic/synth_anlogic.cc b/techlibs/anlogic/synth_anlogic.cc
index 68f4399d4..fab199fd7 100644
--- a/techlibs/anlogic/synth_anlogic.cc
+++ b/techlibs/anlogic/synth_anlogic.cc
@@ -170,6 +170,7 @@ struct SynthAnlogicPass : public ScriptPass
 		{
 			run("dffsr2dff");
 			run("techmap -D NO_LUT -map +/anlogic/cells_map.v");
+			run("dffinit -strinit SET RESET -ff AL_MAP_SEQ q REGSET -noreinit");
 			run("opt_expr -mux_undef");
 			run("simplemap");
 		}
diff --git a/techlibs/common/Makefile.inc b/techlibs/common/Makefile.inc
index 70074f653..0e05620bc 100644
--- a/techlibs/common/Makefile.inc
+++ b/techlibs/common/Makefile.inc
@@ -26,5 +26,5 @@ $(eval $(call add_share_file,share,techlibs/common/pmux2mux.v))
 $(eval $(call add_share_file,share,techlibs/common/adff2dff.v))
 $(eval $(call add_share_file,share,techlibs/common/dff2ff.v))
 $(eval $(call add_share_file,share,techlibs/common/gate2lut.v))
+$(eval $(call add_share_file,share,techlibs/common/cmp2lut.v))
 $(eval $(call add_share_file,share,techlibs/common/cells.lib))
-
diff --git a/techlibs/common/cmp2lut.v b/techlibs/common/cmp2lut.v
new file mode 100644
index 000000000..8aa1eb957
--- /dev/null
+++ b/techlibs/common/cmp2lut.v
@@ -0,0 +1,105 @@
+// Certain arithmetic operations between a signal of width n and a constant can be directly mapped
+// to a single k-LUT (where n <= k). This is preferable to normal alumacc techmapping process
+// because for many targets, arithmetic techmapping creates hard logic (such as carry cells) which often
+// cannot be optimized further.
+//
+// TODO: Currently, only comparisons with 1-bit output are mapped. Potentially, all arithmetic cells
+// with n <= k inputs should be techmapped in this way, because this shortens the critical path
+// from n to 1 by avoiding carry chains.
+
+(* techmap_celltype = "$eq $ne $lt $le $gt $ge" *)
+module _90_lut_cmp_ (A, B, Y);
+
+parameter A_SIGNED = 0;
+parameter B_SIGNED = 0;
+parameter A_WIDTH = 0;
+parameter B_WIDTH = 0;
+parameter Y_WIDTH = 0;
+
+input [A_WIDTH-1:0] A;
+input [B_WIDTH-1:0] B;
+output [Y_WIDTH-1:0] Y;
+
+parameter _TECHMAP_CELLTYPE_ = "";
+
+parameter _TECHMAP_CONSTMSK_A_ = 0;
+parameter _TECHMAP_CONSTVAL_A_ = 0;
+parameter _TECHMAP_CONSTMSK_B_ = 0;
+parameter _TECHMAP_CONSTVAL_B_ = 0;
+
+function automatic integer gen_lut;
+	input integer width;
+	input integer operation;
+	input integer swap;
+	input integer sign;
+	input integer operand;
+	integer n, i_var, i_cst, lhs, rhs, o_bit;
+	begin
+		gen_lut = width'b0;
+		for (n = 0; n < (1 << width); n++) begin
+			if (sign)
+				i_var = n[width-1:0];
+			else
+				i_var = n;
+			i_cst = operand;
+			if (swap) begin
+				lhs = i_cst;
+				rhs = i_var;
+			end else begin
+				lhs = i_var;
+				rhs = i_cst;
+			end
+			if (operation == 0)
+				o_bit = (lhs <  rhs);
+			if (operation == 1)
+				o_bit = (lhs <= rhs);
+			if (operation == 2)
+				o_bit = (lhs >  rhs);
+			if (operation == 3)
+				o_bit = (lhs >= rhs);
+			if (operation == 4)
+				o_bit = (lhs == rhs);
+			if (operation == 5)
+				o_bit = (lhs != rhs);
+			gen_lut = gen_lut | (o_bit << n);
+		end
+	end
+endfunction
+
+generate
+	if (_TECHMAP_CELLTYPE_ == "$lt")
+		localparam operation = 0;
+	if (_TECHMAP_CELLTYPE_ == "$le")
+		localparam operation = 1;
+	if (_TECHMAP_CELLTYPE_ == "$gt")
+		localparam operation = 2;
+	if (_TECHMAP_CELLTYPE_ == "$ge")
+		localparam operation = 3;
+	if (_TECHMAP_CELLTYPE_ == "$eq")
+		localparam operation = 4;
+	if (_TECHMAP_CELLTYPE_ == "$ne")
+		localparam operation = 5;
+
+	if (A_WIDTH > `LUT_WIDTH || B_WIDTH > `LUT_WIDTH || Y_WIDTH != 1)
+		wire _TECHMAP_FAIL_ = 1;
+	else if (&_TECHMAP_CONSTMSK_B_)
+		\$lut #(
+			.WIDTH(A_WIDTH),
+			.LUT({ gen_lut(A_WIDTH, operation, 0, A_SIGNED && B_SIGNED, _TECHMAP_CONSTVAL_B_) })
+		) _TECHMAP_REPLACE_ (
+			.A(A),
+			.Y(Y)
+		);
+	else if (&_TECHMAP_CONSTMSK_A_)
+		\$lut #(
+			.WIDTH(B_WIDTH),
+			.LUT({ gen_lut(B_WIDTH, operation, 1, A_SIGNED && B_SIGNED, _TECHMAP_CONSTVAL_A_) })
+		) _TECHMAP_REPLACE_ (
+			.A(B),
+			.Y(Y)
+		);
+	else
+		wire _TECHMAP_FAIL_ = 1;
+endgenerate
+
+endmodule
diff --git a/techlibs/common/synth.cc b/techlibs/common/synth.cc
index d9565131e..ccfa76e02 100644
--- a/techlibs/common/synth.cc
+++ b/techlibs/common/synth.cc
@@ -51,6 +51,9 @@ struct SynthPass : public ScriptPass
 		log("    -encfile <file>\n");
 		log("        passed to 'fsm_recode' via 'fsm'\n");
 		log("\n");
+		log("    -lut <k>\n");
+		log("        perform synthesis for a k-LUT architecture.\n");
+		log("\n");
 		log("    -nofsm\n");
 		log("        do not run FSM optimization\n");
 		log("\n");
@@ -80,6 +83,7 @@ struct SynthPass : public ScriptPass
 
 	string top_module, fsm_opts, memory_opts;
 	bool autotop, flatten, noalumacc, nofsm, noabc, noshare;
+	int lut;
 
 	void clear_flags() YS_OVERRIDE
 	{
@@ -89,6 +93,7 @@ struct SynthPass : public ScriptPass
 
 		autotop = false;
 		flatten = false;
+		lut = 0;
 		noalumacc = false;
 		nofsm = false;
 		noabc = false;
@@ -130,6 +135,10 @@ struct SynthPass : public ScriptPass
 				flatten = true;
 				continue;
 			}
+			if (args[argidx] == "-lut") {
+				lut = atoi(args[++argidx].c_str());
+				continue;
+			}
 			if (args[argidx] == "-nofsm") {
 				nofsm = true;
 				continue;
@@ -186,19 +195,23 @@ struct SynthPass : public ScriptPass
 		{
 			run("proc");
 			if (help_mode || flatten)
-				run("flatten", "(if -flatten)");
+				run("flatten", "  (if -flatten)");
 			run("opt_expr");
 			run("opt_clean");
 			run("check");
 			run("opt");
 			run("wreduce");
+			if (help_mode)
+				run("techmap -map +/cmp2lut.v", " (if -lut)");
+			else
+				run(stringf("techmap -map +/cmp2lut.v -D LUT_WIDTH=%d", lut));
 			if (!noalumacc)
-				run("alumacc");
+				run("alumacc", "  (unless -noalumacc)");
 			if (!noshare)
-				run("share");
+				run("share", "    (unless -noshare)");
 			run("opt");
 			if (!nofsm)
-				run("fsm" + fsm_opts);
+				run("fsm" + fsm_opts, "      (unless -nofsm)");
 			run("opt -fast");
 			run("memory -nomap" + memory_opts);
 			run("opt_clean");
@@ -210,12 +223,33 @@ struct SynthPass : public ScriptPass
 			run("memory_map");
 			run("opt -full");
 			run("techmap");
+			if (help_mode)
+			{
+				run("techmap -map +/gate2lut.v", "(if -noabc and -lut)");
+				run("clean; opt_lut", "           (if -noabc and -lut)");
+			}
+			else if (noabc && lut)
+			{
+				run(stringf("techmap -map +/gate2lut.v -D LUT_WIDTH=%d", lut));
+				run("clean; opt_lut");
+			}
 			run("opt -fast");
 
 			if (!noabc) {
 		#ifdef YOSYS_ENABLE_ABC
-				run("abc -fast");
-				run("opt -fast");
+				if (help_mode)
+				{
+					run("abc -fast", "       (unless -noabc, unless -lut)");
+					run("abc -fast -lut k", "(unless -noabc, if -lut)");
+				}
+				else
+				{
+					if (lut)
+						run(stringf("abc -fast -lut %d", lut));
+					else
+						run("abc -fast");
+				}
+				run("opt -fast", "       (unless -noabc)");
 		#endif
 			}
 		}
diff --git a/techlibs/ice40/synth_ice40.cc b/techlibs/ice40/synth_ice40.cc
index 626f6d381..f900453e8 100644
--- a/techlibs/ice40/synth_ice40.cc
+++ b/techlibs/ice40/synth_ice40.cc
@@ -226,7 +226,7 @@ struct SynthIce40Pass : public ScriptPass
 
 		if (check_label("coarse"))
 		{
-			run("synth -run coarse");
+			run("synth -lut 4 -run coarse");
 		}
 
 		if (!nobram && check_label("bram", "(skip if -nobram)"))