Try splitting $shiftx with Y_WIDTH > 1 into Y_WIDTH = 1

1 files changed, 38 insertions, 31 deletions

diff --git a/techlibs/xilinx/cells_map.v b/techlibs/xilinx/cells_map.v
index 758d2ade3..253678028 100644
--- a/techlibs/xilinx/cells_map.v
+++ b/techlibs/xilinx/cells_map.v
@@ -32,58 +32,65 @@ module \$shiftx (A, B, Y);
   parameter [B_WIDTH-1:0] _TECHMAP_CONSTVAL_B_ = 0;
 
   generate
-    genvar i;
-    if (B_WIDTH < 3) begin
-      wire _TECHMAP_FAIL_;
-      assign _TECHMAP_FAIL_ = 1;
+    genvar i, j;
+    if (B_SIGNED) begin
+      if (_TECHMAP_CONSTMSK_B_[B_WIDTH-1] && _TECHMAP_CONSTVAL_B_[B_WIDTH-1] == 1'b0)
+        // Optimisation to remove B_SIGNED if sign bit of B is constant-0
+        \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(0), .A_WIDTH(A_WIDTH), .B_WIDTH(B_WIDTH-1), .Y_WIDTH(Y_WIDTH)) _TECHMAP_REPLACE_ (.A(A), .B(B[B_WIDTH-2:0]), .Y(Y));
+      else
+        wire _TECHMAP_FAIL_ = 1;
     end
-    // Optimisation to remove B_SIGNED if sign bit of B is constant-0
-    else if (B_SIGNED && _TECHMAP_CONSTMSK_B_[B_WIDTH-1] && _TECHMAP_CONSTVAL_B_[B_WIDTH-1] == 1'b0) begin
-      \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(0), .A_WIDTH(A_WIDTH), .B_WIDTH(B_WIDTH-1), .Y_WIDTH(Y_WIDTH)) _TECHMAP_REPLACE_ (.A(A), .B(B[B_WIDTH-2:0]), .Y(Y));
+    else if (Y_WIDTH > 1) begin
+      for (i = 0; i < Y_WIDTH; i++) begin
+        wire [A_WIDTH/Y_WIDTH-1:0] A_i;
+        for (j = 0; j < A_WIDTH/Y_WIDTH; j++)
+          assign A_i[j] = A[i*Y_WIDTH+j];
+        wire [$clog2(A_WIDTH/Y_WIDTH)-1:0] B_i = B/Y_WIDTH;
+        \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(A_WIDTH/Y_WIDTH), .B_WIDTH($clog2(A_WIDTH/Y_WIDTH)), .Y_WIDTH(1)) bitblast (.A(A_i), .B(B_i), .Y(Y[i]));
+      end
+    end
+    else if (B_WIDTH < 3) begin
+      wire _TECHMAP_FAIL_ = 1;
     end
     else if (B_WIDTH == 3) begin
-      localparam a_width0 = Y_WIDTH * (2 ** (B_WIDTH-1));
+      localparam a_width0 = 2 ** 2;
       localparam a_widthN = A_WIDTH - a_width0;
-      wire [Y_WIDTH-1:0] T0, T1;
+      wire T0, T1;
       \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_width0), .B_WIDTH(B_WIDTH-1),        .Y_WIDTH(Y_WIDTH)) fpga_shiftx      (.A(A[a_width0-1:0]),       .B(B[B_WIDTH-2:0]), .Y(T0));
       \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_widthN), .B_WIDTH($clog2(a_widthN)), .Y_WIDTH(Y_WIDTH)) fpga_shiftx_last (.A(A[A_WIDTH-1:a_width0]), .B(B[$clog2(a_widthN)-1:0]), .Y(T1));
-      //MUXF7 fpga_mux[Y_WIDTH-1:0] (.I0(T0), .I1(T1), .S(B[B_WIDTH-1]), .O(Y));
-      for (i = 0; i < Y_WIDTH; i++)
-        MUXF7 fpga_mux (.I0(T0[i]), .I1(T1[i]), .S(B[B_WIDTH-1]), .O(Y[i]));
+      MUXF7 fpga_mux (.I0(T0[i]), .I1(T1[i]), .S(B[B_WIDTH-1]), .O(Y[i]));
     end
     else if (B_WIDTH == 4) begin
-      localparam a_width0 = Y_WIDTH * (2 ** (B_WIDTH-2));
+      localparam a_width0 = 2 ** 3;
       localparam num_mux8 = A_WIDTH / a_width0;
       localparam a_widthN = A_WIDTH - num_mux8*a_width0;
-      wire [Y_WIDTH*B_WIDTH-1:0] T;
-      wire [Y_WIDTH-1:0] T0, T1;
+      wire [B_WIDTH-1:0] T;
+      wire T0, T1;
       for (i = 0; i < B_WIDTH; i++)
-	if (i < num_mux8)
-          \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_width0), .B_WIDTH(B_WIDTH-2),        .Y_WIDTH(Y_WIDTH)) fpga_shiftx      (.A(A[(i+1)*a_width0-1:i*a_width0]), .B(B[B_WIDTH-3:0]),          .Y(T[(i+1)*Y_WIDTH-1:i*Y_WIDTH]));
+        if (i < num_mux8)
+          \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_width0), .B_WIDTH(B_WIDTH-2),        .Y_WIDTH(Y_WIDTH)) fpga_shiftx      (.A(A[(i+1)*a_width0-1:i*a_width0]), .B(B[B_WIDTH-3:0]),          .Y(T[i]));
         else if (i == num_mux8 && a_widthN > 0)
-	  \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_widthN), .B_WIDTH($clog2(a_widthN)), .Y_WIDTH(Y_WIDTH)) fpga_shiftx_last (.A(A[A_WIDTH-1:i*a_width0]),        .B(B[$clog2(a_widthN)-1:0]), .Y(T[(i+1)*Y_WIDTH-1:i*Y_WIDTH]));
+          \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_widthN), .B_WIDTH($clog2(a_widthN)), .Y_WIDTH(Y_WIDTH)) fpga_shiftx_last (.A(A[A_WIDTH-1:i*a_width0]),        .B(B[$clog2(a_widthN)-1:0]), .Y(T[i]));
         else
-	  assign T[(i+1)*Y_WIDTH-1:i*Y_WIDTH] = {Y_WIDTH{1'bx}};
-      for (i = 0; i < Y_WIDTH; i++) begin
-        MUXF7 fpga_mux_0 (.I0(T[i*B_WIDTH+0]), .I1(T[i*B_WIDTH+1]), .S(B[B_WIDTH-2]), .O(T0[i]));
-        MUXF7 fpga_mux_1 (.I0(T[i*B_WIDTH+2]), .I1(T[i*B_WIDTH+3]), .S(B[B_WIDTH-2]), .O(T1[i]));
-        MUXF8 fpga_mux_2 (.I0(T0[i]), .I1(T1[i]), .S(B[B_WIDTH-1]), .O(Y[i]));
-      end
+          assign T[i] = 1'bx;
+      MUXF7 fpga_mux_0 (.I0(T[0]), .I1(T[1]), .S(B[B_WIDTH-2]), .O(T0));
+      MUXF7 fpga_mux_1 (.I0(T[2]), .I1(T[3]), .S(B[B_WIDTH-2]), .O(T1));
+      MUXF8 fpga_mux_2 (.I0(T0),   .I1(T1),   .S(B[B_WIDTH-1]), .O(Y));
     end
     else begin
-      localparam a_width0 = Y_WIDTH * (2 ** 4);
+      localparam a_width0 = 2 ** 4;
       localparam num_mux16 = A_WIDTH / a_width0;
       localparam a_widthN = A_WIDTH - num_mux16*a_width0;
-      wire [Y_WIDTH*(2**(B_WIDTH-4))-1:0] T;
+      wire [(2**(B_WIDTH-4))-1:0] T;
       for (i = 0; i < 2 ** (B_WIDTH-4); i++)
-	if (i < num_mux16)
-          \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_width0), .B_WIDTH(4),                .Y_WIDTH(Y_WIDTH)) fpga_shiftx      (.A(A[(i+1)*a_width0-1:i*a_width0]), .B(B[4-1:0]),                .Y(T[(i+1)*Y_WIDTH-1:i*Y_WIDTH]));
+        if (i < num_mux16)
+          \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_width0), .B_WIDTH(4),                .Y_WIDTH(Y_WIDTH)) fpga_shiftx      (.A(A[(i+1)*a_width0-1:i*a_width0]), .B(B[4-1:0]),                .Y(T[i]));
         else if (i == num_mux16 && a_widthN > 0) begin
-	  \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_widthN), .B_WIDTH($clog2(a_widthN)), .Y_WIDTH(Y_WIDTH)) fpga_shiftx_last (.A(A[A_WIDTH-1:i*a_width0]),        .B(B[$clog2(a_widthN)-1:0]), .Y(T[(i+1)*Y_WIDTH-1:i*Y_WIDTH]));
+          \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(a_widthN), .B_WIDTH($clog2(a_widthN)), .Y_WIDTH(Y_WIDTH)) fpga_shiftx_last (.A(A[A_WIDTH-1:i*a_width0]),        .B(B[$clog2(a_widthN)-1:0]), .Y(T[i]));
         end
         else
-	  assign T[(i+1)*Y_WIDTH-1:i*Y_WIDTH] = {Y_WIDTH{1'bx}};
-      \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(Y_WIDTH*(2**(B_WIDTH-4))), .B_WIDTH(B_WIDTH-4), .Y_WIDTH(Y_WIDTH)) fpga_shiftx (.A(T), .B(B[B_WIDTH-1:4]), .Y(Y));
+          assign T[i] = 1'bx;
+      \$shiftx  #(.A_SIGNED(A_SIGNED), .B_SIGNED(B_SIGNED), .A_WIDTH(2**(B_WIDTH-4)), .B_WIDTH(B_WIDTH-4), .Y_WIDTH(Y_WIDTH)) fpga_shiftx (.A(T), .B(B[B_WIDTH-1:4]), .Y(Y));
     end
   endgenerate
 endmodule