Merge remote-tracking branch 'origin/master' into xaig_dff

12 files changed, 228 insertions, 98 deletions

diff --git a/techlibs/xilinx/Makefile.inc b/techlibs/xilinx/Makefile.inc
index 12ec20053..860fcd88c 100644
--- a/techlibs/xilinx/Makefile.inc
+++ b/techlibs/xilinx/Makefile.inc
@@ -30,9 +30,11 @@ $(eval $(call add_share_file,share/xilinx,techlibs/xilinx/drams_map.v))
 $(eval $(call add_share_file,share/xilinx,techlibs/xilinx/arith_map.v))
 $(eval $(call add_share_file,share/xilinx,techlibs/xilinx/ff_map.v))
 $(eval $(call add_share_file,share/xilinx,techlibs/xilinx/lut_map.v))
+
 $(eval $(call add_share_file,share/xilinx,techlibs/xilinx/abc_ff.v))
 $(eval $(call add_share_file,share/xilinx,techlibs/xilinx/abc_xc7.box))
 $(eval $(call add_share_file,share/xilinx,techlibs/xilinx/abc_xc7.lut))
+$(eval $(call add_share_file,share/xilinx,techlibs/xilinx/abc_xc7_nowide.lut))
 
 $(eval $(call add_gen_share_file,share/xilinx,techlibs/xilinx/brams_init_36.vh))
 $(eval $(call add_gen_share_file,share/xilinx,techlibs/xilinx/brams_init_32.vh))
diff --git a/techlibs/xilinx/abc_xc7.box b/techlibs/xilinx/abc_xc7.box
index 4caf69320..bb9258e78 100644
--- a/techlibs/xilinx/abc_xc7.box
+++ b/techlibs/xilinx/abc_xc7.box
@@ -1,5 +1,8 @@
 # Max delays from https://github.com/SymbiFlow/prjxray-db/blob/34ea6eb08a63d21ec16264ad37a0a7b142ff6031/artix7/timings/CLBLL_L.sdf
 
+# NB: Inputs/Outputs must be ordered alphabetically
+#     (with exceptions for carry in/out)
+
 # F7BMUX slower than F7AMUX
 # Inputs: I0 I1 S0
 # Outputs: O
@@ -12,34 +15,42 @@ MUXF8 2 1 3 1
 104 94 273
 
 # CARRY4 + CARRY4_[ABCD]X
-# Inputs: S0 S1 S2 S3 CYINIT DI0 DI1 DI2 DI3 CI
+# Inputs: CYINIT DI0 DI1 DI2 DI3 S0 S1 S2 S3 CI
 # Outputs:  O0 O1 O2 O3 CO0 CO1 CO2 CO3
-#   (NB: carry chain input/output must be last input/output,
-#        swapped with what normally would have been the last
-#        output, here: CI <-> S, CO <-> O
+#   (NB: carry chain input/output must be last
+#        input/output and the entire bus has been
+#        moved there overriding the otherwise
+#        alphabetical ordering)
 CARRY4 3 1 10 8
-223 -   -   -   482 -   -   -   -   222
-400 205 -   -   598 407 -   -   -   334
-523 558 226 -   584 556 537 -   -   239
-582 618 330 227 642 615 596 438 -   313
-340 -   -   -   536 379 -   -   -   271
-433 469 -   -   494 465 445 -   -   157
-512 548 292 -   592 540 520 356 -   228
-508 528 378 380 580 526 507 398 385 114
+482 -   -   -   -   223 -   -   -   222
+598 407 -   -   -   400 205 -   -   334
+584 556 537 -   -   523 558 226 -   239
+642 615 596 438 -   582 618 330 227 313
+536 379 -   -   -   340 -   -   -   271
+494 465 445 -   -   433 469 -   -   157
+592 540 520 356 -   512 548 292 -   228
+580 526 507 398 385 508 528 378 380 114
+
+# SLICEM/A6LUT
+# Inputs: A0 A1 A2 A3 A4 D DPRA0 DPRA1 DPRA2 DPRA3 DPRA4 WCLK WE
+# Outputs: DPO SPO
+RAM32X1D 4 0 13 2
+-   -   -   -   -   - 631 472 407 238 127 - -
+631 472 407 238 127 - -   -   -   -   -   - -
 
 # SLICEM/A6LUT
 # Inputs: A0 A1 A2 A3 A4 A5 D DPRA0 DPRA1 DPRA2 DPRA3 DPRA4 DPRA5 WCLK WE
 # Outputs: DPO SPO
-RAM64X1D 4 0 15 2
--   -   -   -   -   -   - 124 124 124 124 124 124 - -
-124 124 124 124 124 124 - -   -   -   -   -   124 - -
+RAM64X1D 5 0 15 2
+-   -   -   -   -   -   - 642 631 472 407 238 127 - -
+642 631 472 407 238 127 - -   -   -   -   -   -   - -
 
 # SLICEM/A6LUT + F7[AB]MUX
 # Inputs: A0 A1 A2 A3 A4 A5 A6 D DPRA0 DPRA1 DPRA2 DPRA3 DPRA4 DPRA5 DPRA6 WCLK WE
 # Outputs: DPO SPO
-RAM128X1D 5 0 17 2
--   -   -   -   -   -   -   - 314 314 314 314 314 314 292 - -
-347 347 347 347 347 347 296 - -   -   -   -   -   -   -   - -
+RAM128X1D 6 0 17 2
+-    -    -   -   -   -   -   - 1009 998 839 774 605 494 450 - -
+1047 1036 877 812 643 532 478 - -    -   -   -   -   -   -   - -
 
 # Inputs: C CE D R \$pastQ
 # Outputs: Q
diff --git a/techlibs/xilinx/abc_xc7.lut b/techlibs/xilinx/abc_xc7.lut
index f69a923d0..bcbdec127 100644
--- a/techlibs/xilinx/abc_xc7.lut
+++ b/techlibs/xilinx/abc_xc7.lut
@@ -8,7 +8,7 @@
 4	3	127 238 407 472
 5	3	127 238 407 472 631
 6	5	127 238 407 472 631 642
-		# F7AMUX.S+F7BMUX.S + AOUTMUX+COUTMUX / 2
+		# (F7[AB]MUX.S + [AC]OUTMUX) / 2
 7	10	464 513 624 793 858 1017 1028
 		# F8MUX.S+BOUTMUX
 		    # F8MUX.I0+F7MUX.S+BOUTMUX
diff --git a/techlibs/xilinx/abc_xc7_nowide.lut b/techlibs/xilinx/abc_xc7_nowide.lut
new file mode 100644
index 000000000..fab48c879
--- /dev/null
+++ b/techlibs/xilinx/abc_xc7_nowide.lut
@@ -0,0 +1,10 @@
+# Max delays from https://github.com/SymbiFlow/prjxray-db/blob/82bf5f158cd8e9a11ac4d04f1aeef48ed1a528a5/artix7/timings/CLBLL_L.sdf
+#            and  https://github.com/SymbiFlow/prjxray-db/blob/82bf5f158cd8e9a11ac4d04f1aeef48ed1a528a5/artix7/tile_type_CLBLL_L.json
+
+# K	area	delay
+1 	1	127
+2	2	127 238
+3	3	127 238 407
+4	3	127 238 407 472
+5	3	127 238 407 472 631
+6	5	127 238 407 472 631 642
diff --git a/techlibs/xilinx/cells_map.v b/techlibs/xilinx/cells_map.v
index b5114758c..9a316fc96 100644
--- a/techlibs/xilinx/cells_map.v
+++ b/techlibs/xilinx/cells_map.v
@@ -20,16 +20,14 @@
 
 // Convert negative-polarity reset to positive-polarity
 (* techmap_celltype = "$_DFF_NN0_" *)
-module _90_dff_nn0_to_np0(input D, C, R, output Q); \$_DFF_NP0_  _TECHMAP_REPLACE_ (.D(D), .Q(Q), .C(C), .R(~R)); endmodule
+module _90_dff_nn0_to_np0 (input D, C, R, output Q); \$_DFF_NP0_  _TECHMAP_REPLACE_ (.D(D), .Q(Q), .C(C), .R(~R)); endmodule
 (* techmap_celltype = "$_DFF_PN0_" *)
-module _90_dff_pn0_to_pp0(input D, C, R, output Q); \$_DFF_PP0_  _TECHMAP_REPLACE_ (.D(D), .Q(Q), .C(C), .R(~R)); endmodule
-
+module _90_dff_pn0_to_pp0 (input D, C, R, output Q); \$_DFF_PP0_  _TECHMAP_REPLACE_ (.D(D), .Q(Q), .C(C), .R(~R)); endmodule
 (* techmap_celltype = "$_DFF_NN1_" *)
 module _90_dff_nn1_to_np1 (input D, C, R, output Q); \$_DFF_NP1   _TECHMAP_REPLACE_ (.D(D), .Q(Q), .C(C), .R(~R)); endmodule
 (* techmap_celltype = "$_DFF_PN1_" *)
 module _90_dff_pn1_to_pp1 (input D, C, R, output Q); \$_DFF_PP1   _TECHMAP_REPLACE_ (.D(D), .Q(Q), .C(C), .R(~R)); endmodule
 
-
 module \$__SHREG_ (input C, input D, input E, output Q);
   parameter DEPTH = 0;
   parameter [DEPTH-1:0] INIT = 0;
diff --git a/techlibs/xilinx/cells_sim.v b/techlibs/xilinx/cells_sim.v
index bf7a0ed44..5a148be01 100644
--- a/techlibs/xilinx/cells_sim.v
+++ b/techlibs/xilinx/cells_sim.v
@@ -173,8 +173,8 @@ module XORCY(output O, input CI, LI);
   assign O = CI ^ LI;
 endmodule
 
-(* abc_box_id = 3, abc_carry, lib_whitebox *)
-module CARRY4((* abc_carry_out *) output [3:0] CO, output [3:0] O, (* abc_carry_in *) input CI, input CYINIT, input [3:0] DI, S);
+(* abc_box_id = 3, abc_carry="CI,CO", lib_whitebox *)
+module CARRY4(output [3:0] CO, O, input CI, CYINIT, input [3:0] DI, S);
   assign O = S ^ {CO[2:0], CI | CYINIT};
   assign CO[0] = S[0] ? CI | CYINIT : DI[0];
   assign CO[1] = S[1] ? CO[0] : DI[1];
@@ -281,7 +281,25 @@ module FDPE_1 (output reg Q, input C, CE, D, PRE);
   always @(negedge C, posedge PRE) if (PRE) Q <= 1'b1; else if (CE) Q <= D;
 endmodule
 
-//(* abc_box_id = 4 /*, lib_whitebox*/ *)
+(* abc_box_id = 4, abc_scc_break="D,WE" *)
+module RAM32X1D (
+  output DPO, SPO,
+  input  D, WCLK, WE,
+  input  A0, A1, A2, A3, A4,
+  input  DPRA0, DPRA1, DPRA2, DPRA3, DPRA4
+);
+  parameter INIT = 32'h0;
+  parameter IS_WCLK_INVERTED = 1'b0;
+  wire [4:0] a = {A4, A3, A2, A1, A0};
+  wire [4:0] dpra = {DPRA4, DPRA3, DPRA2, DPRA1, DPRA0};
+  reg [31:0] mem = INIT;
+  assign SPO = mem[a];
+  assign DPO = mem[dpra];
+  wire clk = WCLK ^ IS_WCLK_INVERTED;
+  always @(posedge clk) if (WE) mem[a] <= D;
+endmodule
+
+(* abc_box_id = 5, abc_scc_break="D,WE" *)
 module RAM64X1D (
   output DPO, SPO,
   input  D, WCLK, WE,
@@ -299,7 +317,7 @@ module RAM64X1D (
   always @(posedge clk) if (WE) mem[a] <= D;
 endmodule
 
-//(* abc_box_id = 5 /*, lib_whitebox*/ *)
+(* abc_box_id = 6, abc_scc_break="D,WE" *)
 module RAM128X1D (
   output       DPO, SPO,
   input        D, WCLK, WE,
diff --git a/techlibs/xilinx/cells_xtra.sh b/techlibs/xilinx/cells_xtra.sh
index 2b384f405..53b528820 100644
--- a/techlibs/xilinx/cells_xtra.sh
+++ b/techlibs/xilinx/cells_xtra.sh
@@ -120,7 +120,7 @@ function xtract_cell_decl()
 	xtract_cell_decl RAM128X1S
 	xtract_cell_decl RAM256X1S
 	xtract_cell_decl RAM32M
-	xtract_cell_decl RAM32X1D
+	#xtract_cell_decl RAM32X1D
 	xtract_cell_decl RAM32X1S
 	xtract_cell_decl RAM32X1S_1
 	xtract_cell_decl RAM32X2S
diff --git a/techlibs/xilinx/cells_xtra.v b/techlibs/xilinx/cells_xtra.v
index 0ec3d0df0..15fa1b63a 100644
--- a/techlibs/xilinx/cells_xtra.v
+++ b/techlibs/xilinx/cells_xtra.v
@@ -3694,13 +3694,6 @@ module RAM32M (...);
     input WE;
 endmodule
 
-module RAM32X1D (...);
-    parameter [31:0] INIT = 32'h00000000;
-    parameter [0:0] IS_WCLK_INVERTED = 1'b0;
-    output DPO, SPO;
-    input A0, A1, A2, A3, A4, D, DPRA0, DPRA1, DPRA2, DPRA3, DPRA4, WCLK, WE;
-endmodule
-
 module RAM32X1S (...);
     parameter [31:0] INIT = 32'h00000000;
     parameter [0:0] IS_WCLK_INVERTED = 1'b0;
diff --git a/techlibs/xilinx/drams.txt b/techlibs/xilinx/drams.txt
index 91632bcee..2613c206c 100644
--- a/techlibs/xilinx/drams.txt
+++ b/techlibs/xilinx/drams.txt
@@ -1,4 +1,17 @@
 
+bram $__XILINX_RAM32X1D
+  init 1
+  abits 5
+  dbits 1
+  groups 2
+  ports  1 1
+  wrmode 0 1
+  enable 0 1
+  transp 0 0
+  clocks 0 1
+  clkpol 0 2
+endbram
+
 bram $__XILINX_RAM64X1D
   init 1
   abits 6
@@ -25,6 +38,13 @@ bram $__XILINX_RAM128X1D
   clkpol 0 2
 endbram
 
+match $__XILINX_RAM32X1D
+  min bits 3
+  min wports 1
+  make_outreg
+  or_next_if_better
+endmatch
+
 match $__XILINX_RAM64X1D
   min bits 5
   min wports 1
diff --git a/techlibs/xilinx/drams_map.v b/techlibs/xilinx/drams_map.v
index 47476b592..77041ca86 100644
--- a/techlibs/xilinx/drams_map.v
+++ b/techlibs/xilinx/drams_map.v
@@ -1,4 +1,38 @@
 
+module \$__XILINX_RAM32X1D (CLK1, A1ADDR, A1DATA, B1ADDR, B1DATA, B1EN);
+	parameter [31:0] INIT = 32'bx;
+	parameter CLKPOL2 = 1;
+	input CLK1;
+
+	input [4:0] A1ADDR;
+	output A1DATA;
+
+	input [4:0] B1ADDR;
+	input B1DATA;
+	input B1EN;
+
+	RAM32X1D #(
+		.INIT(INIT),
+		.IS_WCLK_INVERTED(!CLKPOL2)
+	) _TECHMAP_REPLACE_ (
+		.DPRA0(A1ADDR[0]),
+		.DPRA1(A1ADDR[1]),
+		.DPRA2(A1ADDR[2]),
+		.DPRA3(A1ADDR[3]),
+		.DPRA4(A1ADDR[4]),
+		.DPO(A1DATA),
+
+		.A0(B1ADDR[0]),
+		.A1(B1ADDR[1]),
+		.A2(B1ADDR[2]),
+		.A3(B1ADDR[3]),
+		.A4(B1ADDR[4]),
+		.D(B1DATA),
+		.WCLK(CLK1),
+		.WE(B1EN)
+	);
+endmodule
+
 module \$__XILINX_RAM64X1D (CLK1, A1ADDR, A1DATA, B1ADDR, B1DATA, B1EN);
 	parameter [63:0] INIT = 64'bx;
 	parameter CLKPOL2 = 1;
diff --git a/techlibs/xilinx/lut_map.v b/techlibs/xilinx/lut_map.v
index d07c59dee..13d3c3268 100644
--- a/techlibs/xilinx/lut_map.v
+++ b/techlibs/xilinx/lut_map.v
@@ -29,61 +29,86 @@ 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
-      LUT1 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
+      LUT1 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
         .I0(A[0]));
     end else
     if (WIDTH == 2) begin
-      LUT2 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[0]), .I1(A[1]));
+      LUT2 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[1]), .I1(A[0]));
     end else
     if (WIDTH == 3) begin
-      LUT3 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[0]), .I1(A[1]), .I2(A[2]));
+      LUT3 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[2]), .I1(A[1]), .I2(A[0]));
     end else
     if (WIDTH == 4) begin
-      LUT4 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[0]), .I1(A[1]), .I2(A[2]),
-        .I3(A[3]));
+      LUT4 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[3]), .I1(A[2]), .I2(A[1]),
+        .I3(A[0]));
     end else
     if (WIDTH == 5) begin
-      LUT5 #(.INIT(LUT)) _TECHMAP_REPLACE_ (.O(Y),
-        .I0(A[0]), .I1(A[1]), .I2(A[2]),
-        .I3(A[3]), .I4(A[4]));
+      LUT5 #(.INIT(P_LUT)) _TECHMAP_REPLACE_ (.O(Y),
+        .I0(A[4]), .I1(A[3]), .I2(A[2]),
+        .I3(A[1]), .I4(A[0]));
     end else
     if (WIDTH == 6) begin
-      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]));
+      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]));
     end else
     if (WIDTH == 7) begin
       wire T0, T1;
-      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]));
+      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]));
     end else
     if (WIDTH == 8) begin
       wire T0, T1, T2, T3, T4, T5;
-      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]));
+      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]));
     end else begin
       wire _TECHMAP_FAIL_ = 1;
     end
diff --git a/techlibs/xilinx/synth_xilinx.cc b/techlibs/xilinx/synth_xilinx.cc
index d3f096220..cdc64db1d 100644
--- a/techlibs/xilinx/synth_xilinx.cc
+++ b/techlibs/xilinx/synth_xilinx.cc
@@ -45,8 +45,9 @@ struct SynthXilinxPass : public ScriptPass
 		log("    -top <module>\n");
 		log("        use the specified module as top module\n");
 		log("\n");
-		log("    -arch {xcup|xcu|xc7|xc6s}\n");
+		log("    -family {xcup|xcu|xc7|xc6s}\n");
 		log("        run synthesis for the specified Xilinx architecture\n");
+		log("        generate the synthesis netlist for the specified family.\n");
 		log("        default: xc7\n");
 		log("\n");
 		log("    -edif <file>\n");
@@ -61,9 +62,6 @@ struct SynthXilinxPass : public ScriptPass
 		log("        generate an output netlist (and BLIF file) suitable for VPR\n");
 		log("        (this feature is experimental and incomplete)\n");
 		log("\n");
-		log("    -nocarry\n");
-		log("        disable inference of carry chains\n");
-		log("\n");
 		log("    -nobram\n");
 		log("        disable inference of block rams\n");
 		log("\n");
@@ -73,6 +71,12 @@ struct SynthXilinxPass : public ScriptPass
 		log("    -nosrl\n");
 		log("        disable inference of shift registers\n");
 		log("\n");
+		log("    -nocarry\n");
+		log("        do not use XORCY/MUXCY/CARRY4 cells in output netlist\n");
+		log("\n");
+		log("    -nowidelut\n");
+		log("        do not use MUXF[78] resources to implement LUTs larger than LUT6s\n");
+		log("\n");
 		log("    -run <from_label>:<to_label>\n");
 		log("        only run the commands between the labels (see below). an empty\n");
 		log("        from label is synonymous to 'begin', and empty to label is\n");
@@ -93,15 +97,15 @@ struct SynthXilinxPass : public ScriptPass
 		log("\n");
 	}
 
-	std::string top_opt, edif_file, blif_file, abc, arch;
-	bool flatten, retime, vpr, nocarry, nobram, nodram, nosrl;
+	std::string top_opt, edif_file, blif_file, family;
+	bool flatten, retime, vpr, nobram, nodram, nosrl, nocarry, nowidelut, abc9;
 
 	void clear_flags() YS_OVERRIDE
 	{
 		top_opt = "-auto-top";
 		edif_file.clear();
 		blif_file.clear();
-		abc = "abc";
+		family = "xc7";
 		flatten = false;
 		retime = false;
 		vpr = false;
@@ -109,7 +113,9 @@ struct SynthXilinxPass : public ScriptPass
 		nobram = false;
 		nodram = false;
 		nosrl = false;
-		arch = "xc7";
+		nocarry = false;
+		nowidelut = false;
+		abc9 = false;
 	}
 
 	void execute(std::vector<std::string> args, RTLIL::Design *design) YS_OVERRIDE
@@ -124,8 +130,8 @@ struct SynthXilinxPass : public ScriptPass
 				top_opt = "-top " + args[++argidx];
 				continue;
 			}
-			if (args[argidx] == "-arch" && argidx+1 < args.size()) {
-				arch = args[++argidx];
+			if ((args[argidx] == "-family" || args[argidx] == "-arch") && argidx+1 < args.size()) {
+				family = args[++argidx];
 				continue;
 			}
 			if (args[argidx] == "-edif" && argidx+1 < args.size()) {
@@ -152,6 +158,14 @@ struct SynthXilinxPass : public ScriptPass
 				retime = true;
 				continue;
 			}
+			if (args[argidx] == "-nocarry") {
+				nocarry = true;
+				continue;
+			}
+			if (args[argidx] == "-nowidelut") {
+				nowidelut = true;
+				continue;
+			}
 			if (args[argidx] == "-vpr") {
 				vpr = true;
 				continue;
@@ -173,15 +187,15 @@ struct SynthXilinxPass : public ScriptPass
 				continue;
 			}
 			if (args[argidx] == "-abc9") {
-				abc = "abc9";
+				abc9 = true;
 				continue;
 			}
 			break;
 		}
 		extra_args(args, argidx, design);
 
-		if (arch != "xcup" && arch != "xcu" && arch != "xc7" && arch != "xc6s")
-			log_cmd_error("Invalid Xilinx -arch setting: %s\n", arch.c_str());
+		if (family != "xcup" && family != "xcu" && family != "xc7" && family != "xc6s")
+			log_cmd_error("Invalid Xilinx -family setting: %s\n", family.c_str());
 
 		if (!design->full_selection())
 			log_cmd_error("This command only operates on fully selected designs!\n");
@@ -225,11 +239,6 @@ struct SynthXilinxPass : public ScriptPass
 			//   so attempt to convert $pmux-es to the former
 			if (!nosrl || help_mode)
 				run("pmux2shiftx", "(skip if '-nosrl')");
-
-			// Run a number of peephole optimisations, including one
-			//   that optimises $mul cells driving $shiftx's B input
-			//   and that aids wide mux analysis
-			run("peepopt");
 		}
 
 		if (check_label("bram", "(skip if '-nobram')")) {
@@ -268,7 +277,7 @@ struct SynthXilinxPass : public ScriptPass
 				techmap_files += " -map +/xilinx/arith_map.v";
 				if (vpr)
 					techmap_files += " -D _EXPLICIT_CARRY";
-				else if (abc == "abc9")
+				else if (abc9)
 					techmap_files += " -D _CLB_CARRY";
 			}
 			run("techmap " + techmap_files);
@@ -276,7 +285,7 @@ struct SynthXilinxPass : public ScriptPass
 		}
 
 		if (check_label("map_cells")) {
-			if (abc == "abc9")
+			if (abc9)
 				run("techmap -map +/techmap.v -map +/xilinx/cells_map.v -D _ABC -map +/xilinx/ff_map.v");
 			else
 				run("techmap -map +/techmap.v -map +/xilinx/cells_map.v");
@@ -284,21 +293,31 @@ struct SynthXilinxPass : public ScriptPass
 		}
 
 		if (check_label("map_luts")) {
-			if (abc == "abc9") {
+			run("opt_expr -mux_undef");
+			if (help_mode)
+				run("abc -luts 2:2,3,6:5[,10,20] [-dff]", "(skip if 'nowidelut', only for '-retime')");
+			else if (abc9) {
+				if (family != "xc7")
+					log_warning("'synth_xilinx -abc9' currently supports '-family xc7' only.\n");
 				run("read_verilog -icells -lib +/xilinx/abc_ff.v");
-				run(abc + " -lut +/xilinx/abc_xc7.lut -box +/xilinx/abc_xc7.box -W " + XC7_WIRE_DELAY + string(retime ? " -retime" : ""));
+				if (nowidelut)
+					run("abc9 -lut +/xilinx/abc_xc7_nowide.lut -box +/xilinx/abc_xc7.box -W " + std::string(XC7_WIRE_DELAY) + string(retime ? " -dff" : ""));
+				else
+					run("abc9 -lut +/xilinx/abc_xc7.lut -box +/xilinx/abc_xc7.box -W " + std::string(XC7_WIRE_DELAY) + string(retime ? " -dff" : ""));
+			}
+			else {
+				if (nowidelut)
+					run("abc -luts 2:2,3,6:5" + string(retime ? " -dff" : ""));
+				else
+					run("abc -luts 2:2,3,6:5,10,20" + string(retime ? " -dff" : ""));
 			}
-			else if (help_mode)
-				run(abc + " -luts 2:2,3,6:5,10,20 [-dff]");
-			else
-				run(abc + " -luts 2:2,3,6:5,10,20" + string(retime ? " -dff" : ""));
 			run("clean");
 
 			// This shregmap call infers fixed length shift registers after abc
 			//   has performed any necessary retiming
 			if (!nosrl || help_mode)
 				run("shregmap -minlen 3 -init -params -enpol any_or_none", "(skip if '-nosrl')");
-			if (abc == "abc9")
+			if (abc9)
 				run("techmap -map +/xilinx/lut_map.v -map +/xilinx/cells_map.v");
 			else
 				run("techmap -map +/xilinx/lut_map.v -map +/xilinx/cells_map.v -map +/xilinx/ff_map.v");