Merge pull request #1660 from YosysHQ/eddie/abc9_unpermute_luts

Unpermute LUT ordering for ice40/ecp5/xilinx

9 files changed, 207 insertions, 149 deletions

diff --git a/frontends/aiger/aigerparse.cc b/frontends/aiger/aigerparse.cc
index 418fd722c..a42569301 100644
--- a/frontends/aiger/aigerparse.cc
+++ b/frontends/aiger/aigerparse.cc
@@ -422,6 +422,8 @@ void AigerReader::parse_xaiger()
 					log_assert(wire);
 					input_sig.append(wire);
 				}
+				// Reverse input order as fastest input is returned first
+				input_sig.reverse();
 				// TODO: Compute LUT mask from AIG in less than O(2 ** input_sig.size())
 				ce.clear();
 				ce.compute_deps(output_sig, input_sig.to_sigbit_pool());
diff --git a/kernel/rtlil.h b/kernel/rtlil.h
index 6251d265d..58c5d9674 100644
--- a/kernel/rtlil.h
+++ b/kernel/rtlil.h
@@ -851,6 +851,8 @@ public:
 
 	RTLIL::SigSpec repeat(int num) const;
 
+	void reverse() { inline_unpack(); std::reverse(bits_.begin(), bits_.end()); }
+
 	bool operator <(const RTLIL::SigSpec &other) const;
 	bool operator ==(const RTLIL::SigSpec &other) const;
 	inline bool operator !=(const RTLIL::SigSpec &other) const { return !(*this == other); }
diff --git a/passes/pmgen/ice40_wrapcarry.cc b/passes/pmgen/ice40_wrapcarry.cc
index d458dce46..0053c8872 100644
--- a/passes/pmgen/ice40_wrapcarry.cc
+++ b/passes/pmgen/ice40_wrapcarry.cc
@@ -127,7 +127,7 @@ struct Ice40WrapCarryPass : public Pass {
 					lut->setParam(ID(WIDTH), 4);
 					lut->setParam(ID(LUT), cell->getParam(ID(LUT)));
 					auto I3 = cell->getPort(cell->getParam(ID(I3_IS_CI)).as_bool() ? ID(CI) : ID(I3));
-					lut->setPort(ID(A), {cell->getPort(ID(I0)), cell->getPort(ID(A)), cell->getPort(ID(B)), I3 });
+					lut->setPort(ID(A), { I3, cell->getPort(ID(B)), cell->getPort(ID(A)), cell->getPort(ID(I0)) });
 					lut->setPort(ID(Y), cell->getPort(ID(O)));
 
 					Const src;
diff --git a/techlibs/ecp5/cells_map.v b/techlibs/ecp5/cells_map.v
index 10e89a3e0..c031703a9 100644
--- a/techlibs/ecp5/cells_map.v
+++ b/techlibs/ecp5/cells_map.v
@@ -73,102 +73,80 @@ module \$lut (A, Y);
     input [WIDTH-1:0] A;
     output Y;
 
-    // Need to swap input ordering, and fix init accordingly,
-    // to match ABC's expectation of LUT inputs in non-decreasing
-    // delay order
-    localparam P_WIDTH = WIDTH < 4 ? 4 : WIDTH;
-    function [P_WIDTH-1:0] permute_index;
-        input [P_WIDTH-1:0] i;
-        integer j;
-        begin
-            permute_index = 0;
-            for (j = 0; j < P_WIDTH; j = j + 1)
-                permute_index[P_WIDTH-1 - j] = i[j];
-        end
-    endfunction
-
-    function [2**P_WIDTH-1:0] permute_init;
-        integer i;
-        begin
-            permute_init = 0;
-            for (i = 0; i < 2**P_WIDTH; i = i + 1)
-                permute_init[i] = LUT[permute_index(i)];
-        end
-    endfunction
-
-    parameter [2**P_WIDTH-1:0] P_LUT = permute_init();
-
     generate
         if (WIDTH == 1) begin
-            LUT4 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.Z(Y),
+            localparam [15:0] INIT = {{8{LUT[1]}}, {8{LUT[0]}}};
+            LUT4 #(.INIT(INIT)) _TECHMAP_REPLACE_ (.Z(Y),
                 .A(1'b0), .B(1'b0), .C(1'b0), .D(A[0]));
         end else
         if (WIDTH == 2) begin
-            LUT4 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.Z(Y),
-                .A(1'b0), .B(1'b0), .C(A[1]), .D(A[0]));
+            localparam [15:0] INIT = {{4{LUT[3]}}, {4{LUT[2]}}, {4{LUT[1]}}, {4{LUT[0]}}};
+            LUT4 #(.INIT(INIT)) _TECHMAP_REPLACE_ (.Z(Y),
+                .A(1'b0), .B(1'b0), .C(A[0]), .D(A[1]));
         end else
         if (WIDTH == 3) begin
-            LUT4 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.Z(Y),
-                .A(1'b0), .B(A[2]), .C(A[1]), .D(A[0]));
+            localparam [15:0] INIT = {{2{LUT[7]}}, {2{LUT[6]}}, {2{LUT[5]}}, {2{LUT[4]}}, {2{LUT[3]}}, {2{LUT[2]}}, {2{LUT[1]}}, {2{LUT[0]}}};
+            LUT4 #(.INIT(INIT)) _TECHMAP_REPLACE_ (.Z(Y),
+                .A(1'b0), .B(A[0]), .C(A[1]), .D(A[2]));
         end else
         if (WIDTH == 4) begin
-            LUT4 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.Z(Y),
-                .A(A[3]), .B(A[2]), .C(A[1]), .D(A[0]));
+            LUT4 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.Z(Y),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
         `ifndef NO_PFUMUX
         end else
         if (WIDTH == 5) begin
             wire f0, f1;
-            LUT4 #(.INIT(P_LUT[15: 0])) lut0 (.Z(f0),
-                .A(A[4]), .B(A[3]), .C(A[2]), .D(A[1]));
-            LUT4 #(.INIT(P_LUT[31:16])) lut1 (.Z(f1),
-                .A(A[4]), .B(A[3]), .C(A[2]), .D(A[1]));
-            PFUMX mux5(.ALUT(f1), .BLUT(f0), .C0(A[0]), .Z(Y));
+            LUT4 #(.INIT(LUT[15: 0])) lut0 (.Z(f0),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            LUT4 #(.INIT(LUT[31:16])) lut1 (.Z(f1),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            PFUMX mux5(.ALUT(f1), .BLUT(f0), .C0(A[4]), .Z(Y));
         end else
         if (WIDTH == 6) begin
             wire f0, f1, f2, f3, g0, g1;
-            LUT4 #(.INIT(P_LUT[15: 0])) lut0 (.Z(f0),
-                .A(A[5]), .B(A[4]), .C(A[3]), .D(A[2]));
-            LUT4 #(.INIT(P_LUT[31:16])) lut1 (.Z(f1),
-                .A(A[5]), .B(A[4]), .C(A[3]), .D(A[2]));
-
-            LUT4 #(.INIT(P_LUT[47:32])) lut2 (.Z(f2),
-                .A(A[5]), .B(A[4]), .C(A[3]), .D(A[2]));
-            LUT4 #(.INIT(P_LUT[63:48])) lut3 (.Z(f3),
-                .A(A[5]), .B(A[4]), .C(A[3]), .D(A[2]));
-
-            PFUMX mux50(.ALUT(f1), .BLUT(f0), .C0(A[1]), .Z(g0));
-            PFUMX mux51(.ALUT(f3), .BLUT(f2), .C0(A[1]), .Z(g1));
-            L6MUX21 mux6 (.D0(g0), .D1(g1), .SD(A[0]), .Z(Y));
+            LUT4 #(.INIT(LUT[15: 0])) lut0 (.Z(f0),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            LUT4 #(.INIT(LUT[31:16])) lut1 (.Z(f1),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+
+            LUT4 #(.INIT(LUT[47:32])) lut2 (.Z(f2),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            LUT4 #(.INIT(LUT[63:48])) lut3 (.Z(f3),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+
+            PFUMX mux50(.ALUT(f1), .BLUT(f0), .C0(A[4]), .Z(g0));
+            PFUMX mux51(.ALUT(f3), .BLUT(f2), .C0(A[4]), .Z(g1));
+            L6MUX21 mux6 (.D0(g0), .D1(g1), .SD(A[5]), .Z(Y));
         end else
         if (WIDTH == 7) begin
             wire f0, f1, f2, f3, f4, f5, f6, f7, g0, g1, g2, g3, h0, h1;
-            LUT4 #(.INIT(P_LUT[15: 0])) lut0 (.Z(f0),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-            LUT4 #(.INIT(P_LUT[31:16])) lut1 (.Z(f1),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-
-            LUT4 #(.INIT(P_LUT[47:32])) lut2 (.Z(f2),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-            LUT4 #(.INIT(P_LUT[63:48])) lut3 (.Z(f3),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-
-            LUT4 #(.INIT(P_LUT[79:64])) lut4 (.Z(f4),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-            LUT4 #(.INIT(P_LUT[95:80])) lut5 (.Z(f5),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-
-            LUT4 #(.INIT(P_LUT[111: 96])) lut6 (.Z(f6),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-            LUT4 #(.INIT(P_LUT[127:112])) lut7 (.Z(f7),
-                .A(A[6]), .B(A[5]), .C(A[4]), .D(A[3]));
-
-            PFUMX mux50(.ALUT(f1), .BLUT(f0), .C0(A[2]), .Z(g0));
-            PFUMX mux51(.ALUT(f3), .BLUT(f2), .C0(A[2]), .Z(g1));
-            PFUMX mux52(.ALUT(f5), .BLUT(f4), .C0(A[2]), .Z(g2));
-            PFUMX mux53(.ALUT(f7), .BLUT(f6), .C0(A[2]), .Z(g3));
-            L6MUX21 mux60 (.D0(g0), .D1(g1), .SD(A[1]), .Z(h0));
-            L6MUX21 mux61 (.D0(g2), .D1(g3), .SD(A[1]), .Z(h1));
-            L6MUX21 mux7  (.D0(h0), .D1(h1), .SD(A[0]), .Z(Y));
+            LUT4 #(.INIT(LUT[15: 0])) lut0 (.Z(f0),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            LUT4 #(.INIT(LUT[31:16])) lut1 (.Z(f1),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+
+            LUT4 #(.INIT(LUT[47:32])) lut2 (.Z(f2),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            LUT4 #(.INIT(LUT[63:48])) lut3 (.Z(f3),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+
+            LUT4 #(.INIT(LUT[79:64])) lut4 (.Z(f4),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            LUT4 #(.INIT(LUT[95:80])) lut5 (.Z(f5),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+
+            LUT4 #(.INIT(LUT[111: 96])) lut6 (.Z(f6),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+            LUT4 #(.INIT(LUT[127:112])) lut7 (.Z(f7),
+                .A(A[0]), .B(A[1]), .C(A[2]), .D(A[3]));
+
+            PFUMX mux50(.ALUT(f1), .BLUT(f0), .C0(A[4]), .Z(g0));
+            PFUMX mux51(.ALUT(f3), .BLUT(f2), .C0(A[4]), .Z(g1));
+            PFUMX mux52(.ALUT(f5), .BLUT(f4), .C0(A[4]), .Z(g2));
+            PFUMX mux53(.ALUT(f7), .BLUT(f6), .C0(A[4]), .Z(g3));
+            L6MUX21 mux60 (.D0(g0), .D1(g1), .SD(A[5]), .Z(h0));
+            L6MUX21 mux61 (.D0(g2), .D1(g3), .SD(A[5]), .Z(h1));
+            L6MUX21 mux7  (.D0(h0), .D1(h1), .SD(A[6]), .Z(Y));
         `endif
         end else begin
             wire _TECHMAP_FAIL_ = 1;
diff --git a/techlibs/ice40/cells_map.v b/techlibs/ice40/cells_map.v
index 759549e30..d5362eb83 100644
--- a/techlibs/ice40/cells_map.v
+++ b/techlibs/ice40/cells_map.v
@@ -42,19 +42,18 @@ module \$lut (A, Y);
         .I0(1'b0), .I1(1'b0), .I2(1'b0), .I3(A[0]));
     end else
     if (WIDTH == 2) begin
-      localparam [15:0] INIT = {{4{LUT[3]}}, {4{LUT[1]}}, {4{LUT[2]}}, {4{LUT[0]}}};
+      localparam [15:0] INIT = {{4{LUT[3]}}, {4{LUT[2]}}, {4{LUT[1]}}, {4{LUT[0]}}};
       SB_LUT4 #(.LUT_INIT(INIT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(1'b0), .I1(1'b0), .I2(A[1]), .I3(A[0]));
+        .I0(1'b0), .I1(1'b0), .I2(A[0]), .I3(A[1]));
     end else
     if (WIDTH == 3) begin
-      localparam [15:0] INIT = {{2{LUT[7]}}, {2{LUT[3]}}, {2{LUT[5]}}, {2{LUT[1]}}, {2{LUT[6]}}, {2{LUT[2]}}, {2{LUT[4]}}, {2{LUT[0]}}};
+      localparam [15:0] INIT = {{2{LUT[7]}}, {2{LUT[6]}}, {2{LUT[5]}}, {2{LUT[4]}}, {2{LUT[3]}}, {2{LUT[2]}}, {2{LUT[1]}}, {2{LUT[0]}}};
       SB_LUT4 #(.LUT_INIT(INIT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(1'b0), .I1(A[2]), .I2(A[1]), .I3(A[0]));
+        .I0(1'b0), .I1(A[0]), .I2(A[1]), .I3(A[2]));
     end else
     if (WIDTH == 4) begin
-      localparam [15:0] INIT = {LUT[15], LUT[7], LUT[11], LUT[3], LUT[13], LUT[5], LUT[9], LUT[1], LUT[14], LUT[6], LUT[10], LUT[2], LUT[12], LUT[4], LUT[8], LUT[0]};
-      SB_LUT4 #(.LUT_INIT(INIT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[3]), .I1(A[2]), .I2(A[1]), .I3(A[0]));
+      SB_LUT4 #(.LUT_INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]), .I3(A[3]));
     end else begin
       wire _TECHMAP_FAIL_ = 1;
     end
diff --git a/techlibs/ice40/ice40_opt.cc b/techlibs/ice40/ice40_opt.cc
index df10a2842..925ab31bb 100644
--- a/techlibs/ice40/ice40_opt.cc
+++ b/techlibs/ice40/ice40_opt.cc
@@ -139,8 +139,8 @@ static void run_ice40_opts(Module *module)
 				log("Optimized $__ICE40_CARRY_WRAPPER cell back to logic (without SB_CARRY) %s.%s: CO=%s\n",
 						log_id(module), log_id(cell), log_signal(replacement_output));
 				cell->type = "$lut";
-				auto I3 = cell->getPort(cell->getParam(ID(I3_IS_CI)).as_bool() ? ID(CI) : ID(I3));
-				cell->setPort("\\A", { cell->getPort("\\I0"), inbit[0], inbit[1], I3 });
+				auto I3 = get_bit_or_zero(cell->getPort(cell->getParam(ID(I3_IS_CI)).as_bool() ? ID(CI) : ID(I3)));
+				cell->setPort("\\A", { I3, inbit[1], inbit[0], get_bit_or_zero(cell->getPort("\\I0")) });
 				cell->setPort("\\Y", cell->getPort("\\O"));
 				cell->unsetPort("\\B");
 				cell->unsetPort("\\CI");
diff --git a/techlibs/xilinx/lut_map.v b/techlibs/xilinx/lut_map.v
index 62d501632..718ec42f1 100644
--- a/techlibs/xilinx/lut_map.v
+++ b/techlibs/xilinx/lut_map.v
@@ -29,90 +29,65 @@ module \$lut (A, Y);
   input [WIDTH-1:0] A;
   output Y;
 
-  // Need to swap input ordering, and fix init accordingly,
-  // to match ABC's expectation of LUT inputs in non-decreasing
-  // delay order
-  function [WIDTH-1:0] permute_index;
-      input [WIDTH-1:0] i;
-      integer j;
-      begin
-          permute_index = 0;
-          for (j = 0; j < WIDTH; j = j + 1)
-              permute_index[WIDTH-1 - j] = i[j];
-      end
-  endfunction
-
-  function [2**WIDTH-1:0] permute_init;
-      input [2**WIDTH-1:0] orig;
-      integer i;
-      begin
-          permute_init = 0;
-          for (i = 0; i < 2**WIDTH; i = i + 1)
-              permute_init[i] = orig[permute_index(i)];
-      end
-  endfunction
-
-  parameter [2**WIDTH-1:0] P_LUT = permute_init(LUT);
-
   generate
     if (WIDTH == 1) begin
-      if (P_LUT == 2'b01) begin
+      if (LUT == 2'b01) begin
         INV _TECHMAP_REPLACE_ (.O(Y), .I(A[0]));
       end else begin
-        LUT1 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        LUT1 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
           .I0(A[0]));
       end
     end else
     if (WIDTH == 2) begin
-      LUT2 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[1]), .I1(A[0]));
+      LUT2 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[0]), .I1(A[1]));
     end else
     if (WIDTH == 3) begin
-      LUT3 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[2]), .I1(A[1]), .I2(A[0]));
+      LUT3 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]));
     end else
     if (WIDTH == 4) begin
-      LUT4 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[3]), .I1(A[2]), .I2(A[1]),
-        .I3(A[0]));
+      LUT4 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]));
     end else
     if (WIDTH == 5) begin
-      LUT5 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[4]), .I1(A[3]), .I2(A[2]),
-        .I3(A[1]), .I4(A[0]));
+      LUT5 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]));
     end else
     if (WIDTH == 6) begin
-      LUT6 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[5]), .I1(A[4]), .I2(A[3]),
-        .I3(A[2]), .I4(A[1]), .I5(A[0]));
+      LUT6 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]), .I5(A[5]));
     end else
     if (WIDTH == 7) begin
       wire T0, T1;
-      LUT6 #(.INIT(P_LUT[63:0])) fpga_lut_0 (.O(T0),
-        .I0(A[6]), .I1(A[5]), .I2(A[4]),
-        .I3(A[3]), .I4(A[2]), .I5(A[1]));
-      LUT6 #(.INIT(P_LUT[127:64])) fpga_lut_1 (.O(T1),
-        .I0(A[6]), .I1(A[5]), .I2(A[4]),
-        .I3(A[3]), .I4(A[2]), .I5(A[1]));
-      MUXF7 fpga_mux_0 (.O(Y), .I0(T0), .I1(T1), .S(A[0]));
+      LUT6 #(.INIT(LUT[63:0])) fpga_lut_0 (.O(T0),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]), .I5(A[5]));
+      LUT6 #(.INIT(LUT[127:64])) fpga_lut_1 (.O(T1),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]), .I5(A[5]));
+      MUXF7 fpga_mux_0 (.O(Y), .I0(T0), .I1(T1), .S(A[6]));
     end else
     if (WIDTH == 8) begin
       wire T0, T1, T2, T3, T4, T5;
-      LUT6 #(.INIT(P_LUT[63:0])) fpga_lut_0 (.O(T0),
-        .I0(A[7]), .I1(A[6]), .I2(A[5]),
-        .I3(A[4]), .I4(A[3]), .I5(A[2]));
-      LUT6 #(.INIT(P_LUT[127:64])) fpga_lut_1 (.O(T1),
-        .I0(A[7]), .I1(A[6]), .I2(A[5]),
-        .I3(A[4]), .I4(A[3]), .I5(A[2]));
-      LUT6 #(.INIT(P_LUT[191:128])) fpga_lut_2 (.O(T2),
-        .I0(A[7]), .I1(A[6]), .I2(A[5]),
-        .I3(A[4]), .I4(A[3]), .I5(A[2]));
-      LUT6 #(.INIT(P_LUT[255:192])) fpga_lut_3 (.O(T3),
-        .I0(A[7]), .I1(A[6]), .I2(A[5]),
-        .I3(A[4]), .I4(A[3]), .I5(A[2]));
-      MUXF7 fpga_mux_0 (.O(T4), .I0(T0), .I1(T1), .S(A[1]));
-      MUXF7 fpga_mux_1 (.O(T5), .I0(T2), .I1(T3), .S(A[1]));
-      MUXF8 fpga_mux_2 (.O(Y), .I0(T4), .I1(T5), .S(A[0]));
+      LUT6 #(.INIT(LUT[63:0])) fpga_lut_0 (.O(T0),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]), .I5(A[5]));
+      LUT6 #(.INIT(LUT[127:64])) fpga_lut_1 (.O(T1),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]), .I5(A[5]));
+      LUT6 #(.INIT(LUT[191:128])) fpga_lut_2 (.O(T2),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]), .I5(A[5]));
+      LUT6 #(.INIT(LUT[255:192])) fpga_lut_3 (.O(T3),
+        .I0(A[0]), .I1(A[1]), .I2(A[2]),
+        .I3(A[3]), .I4(A[4]), .I5(A[5]));
+      MUXF7 fpga_mux_0 (.O(T4), .I0(T0), .I1(T1), .S(A[6]));
+      MUXF7 fpga_mux_1 (.O(T5), .I0(T2), .I1(T3), .S(A[6]));
+      MUXF8 fpga_mux_2 (.O(Y), .I0(T4), .I1(T5), .S(A[7]));
     end else begin
       wire _TECHMAP_FAIL_ = 1;
     end
diff --git a/tests/arch/ice40/bug1597.ys b/tests/arch/ice40/bug1597.ys
new file mode 100644
index 000000000..b7983cfa4
--- /dev/null
+++ b/tests/arch/ice40/bug1597.ys
@@ -0,0 +1,72 @@
+read_verilog <<EOT
+module top (
+    input CLK, PIN_1, PIN_2, PIN_3, PIN_4, PIN_5,
+        PIN_6, PIN_7, PIN_8, PIN_9, PIN_10, PIN_11, PIN_12, PIN_13, PIN_25,
+    output USBPU, PIN_14, PIN_15, PIN_16, PIN_17, PIN_18,
+        PIN_19, PIN_20, PIN_21, PIN_22, PIN_23, PIN_24,
+);
+    assign USBPU = 0;
+
+    wire[5:0] parOut;
+    wire[5:0] chrg;
+
+    assign PIN_14 = parOut[0];
+    assign PIN_15 = parOut[1];
+    assign PIN_16 = parOut[2];
+    assign PIN_17 = parOut[3];
+    assign PIN_18 = parOut[4];
+    assign PIN_19 = parOut[5];
+    assign chrg[0] = PIN_3;
+    assign chrg[1] = PIN_4;
+    assign chrg[2] = PIN_5;
+    assign chrg[3] = PIN_6;
+    assign chrg[4] = PIN_7;
+    assign chrg[5] = PIN_8;
+
+    SSCounter6o sc6(PIN_1, CLK, PIN_2, PIN_9, chrg, parOut);
+
+endmodule
+
+module SSCounter6 (input wire rst, clk, adv, jmp, input wire [5:0] in, output reg[5:0] out);
+    always @(posedge clk, posedge rst)
+        if (rst) out <= 0;
+        else if (adv || jmp) out <= jmp ? in : out + 1;
+endmodule
+
+// Optimized 6 bit counter, it should takes 7 cells.
+/* b[5:1]                       /* b[0]
+1010101010101010 in             1010101010101010 in
+1100110011001100 jmp            1100110011001100 jmp
+1111000011110000 loop           1111000011110000 loop
+1111111100000000 carry          1111111100000000 -
+----------------------          ----------------------
+1000101110111000 out            1000101110001011 out
+   8   B   B   8                   8   B   8   B
+*/
+module SSCounter6o (input wire rst, clk, adv, jmp, input wire [5:0] in, output wire[5:0] out);
+    wire[4:0] co;
+    wire[5:0] lo;
+    wire ien;
+    SB_LUT4 #(.LUT_INIT(16'hFFF0)) lien (ien, 0, 0, adv, jmp);
+    SB_CARRY c0 (co[0], jmp, out[0], 1),
+             c1 (co[1], jmp, out[1], co[0]),
+             c2 (co[2], jmp, out[2], co[1]),
+             c3 (co[3], jmp, out[3], co[2]),
+             c4 (co[4], jmp, out[4], co[3]);
+    SB_DFFER d0 (out[0], clk, ien, rst, lo[0]),
+             d1 (out[1], clk, ien, rst, lo[1]),
+             d2 (out[2], clk, ien, rst, lo[2]),
+             d3 (out[3], clk, ien, rst, lo[3]),
+             d4 (out[4], clk, ien, rst, lo[4]),
+             d5 (out[5], clk, ien, rst, lo[5]);
+    SB_LUT4 #(.LUT_INIT(16'h8B8B)) l0 (lo[0], in[0], jmp, out[0], 0);
+    SB_LUT4 #(.LUT_INIT(16'h8BB8)) l1 (lo[1], in[1], jmp, out[1], co[0]);
+    SB_LUT4 #(.LUT_INIT(16'h8BB8)) l2 (lo[2], in[2], jmp, out[2], co[1]);
+    SB_LUT4 #(.LUT_INIT(16'h8BB8)) l3 (lo[3], in[3], jmp, out[3], co[2]);
+    SB_LUT4 #(.LUT_INIT(16'h8BB8)) l4 (lo[4], in[4], jmp, out[4], co[3]);
+    SB_LUT4 #(.LUT_INIT(16'h8BB8)) l5 (lo[5], in[5], jmp, out[5], co[4]);
+endmodule
+EOT
+hierarchy -top top
+flatten
+equiv_opt -multiclock -map +/ice40/cells_sim.v synth_ice40
diff --git a/tests/arch/ice40/ice40_opt.ys b/tests/arch/ice40/ice40_opt.ys
index 011d98fef..71b68431e 100644
--- a/tests/arch/ice40/ice40_opt.ys
+++ b/tests/arch/ice40/ice40_opt.ys
@@ -86,3 +86,33 @@ select -assert-count 1 t:SB_LUT4
 select -assert-count 1 t:SB_CARRY
 select -assert-count 1 t:SB_CARRY a:keep %i
 select -assert-count 1 t:SB_CARRY c:carry %i
+
+
+design -reset
+read_verilog -icells <<EOT
+module top(input I3, I2, I1, I0, output O, O2);
+	SB_LUT4 #(
+		.LUT_INIT(8'b 1001_0110)
+	) u0 (
+		.I0(I0),
+		.I1(I1),
+		.I2(I2),
+		.I3(),
+		.O(O)
+	);
+	wire CO;
+	\$__ICE40_CARRY_WRAPPER #(
+		.LUT(~8'b 1001_0110),
+		.I3_IS_CI(1'b0)
+	) u1 (
+		.A(1'b0),
+		.B(1'b0),
+		.CI(1'b0),
+		.I0(),
+		.I3(),
+		.CO(CO),
+		.O(O2)
+	);
+endmodule
+EOT
+ice40_opt