diff --git a/ice40/arch.h b/ice40/arch.h
index e67f2aa9..98361132 100644
--- a/ice40/arch.h
+++ b/ice40/arch.h
@@ -141,6 +141,7 @@ static const int max_switch_bits = 5;
 
 NPNR_PACKED_STRUCT(struct SwitchInfoPOD {
     int32_t num_bits;
+    int32_t bel;
     int8_t x, y;
     ConfigBitPOD cbits[max_switch_bits];
 });
@@ -613,7 +614,16 @@ struct Arch : BaseCtx
     bool checkPipAvail(PipId pip) const
     {
         NPNR_ASSERT(pip != PipId());
-        return switches_locked[chip_info->pip_data[pip.index].switch_index] == IdString();
+        int switch_idx = chip_info->pip_data[pip.index].switch_index;
+
+        if (switches_locked[switch_idx] != IdString())
+            return false;
+
+        int bel_idx = chip_info->bits_info->switches[switch_idx].bel;
+        if (bel_idx >= 0 && bel_to_cell[bel_idx] != IdString())
+            return false;
+
+        return true;
     }
 
     IdString getBoundPipNet(PipId pip) const
diff --git a/ice40/bitstream.cc b/ice40/bitstream.cc
index d0d6b205..8b00e878 100644
--- a/ice40/bitstream.cc
+++ b/ice40/bitstream.cc
@@ -221,8 +221,21 @@ std::string tagTileType(TileType &tile)
         NPNR_ASSERT(false);
     }
 }
+
+static BelPin get_one_bel_pin(const Context *ctx, WireId wire)
+{
+    auto pins = ctx->getWireBelPins(wire);
+    NPNR_ASSERT(pins.begin() != pins.end());
+    return *pins.begin();
+}
+
 void write_asc(const Context *ctx, std::ostream &out)
 {
+
+    static const std::vector<int> lut_perm = {
+            4, 14, 15, 5, 6, 16, 17, 7, 3, 13, 12, 2, 1, 11, 10, 0,
+    };
+
     // [y][x][row][col]
     const ChipInfoPOD &ci = *ctx->chip_info;
     const BitstreamInfoPOD &bi = *ci.bits_info;
@@ -262,12 +275,50 @@ void write_asc(const Context *ctx, std::ostream &out)
         if (ctx->pip_to_net[pip.index] != IdString()) {
             const PipInfoPOD &pi = ci.pip_data[pip.index];
             const SwitchInfoPOD &swi = bi.switches[pi.switch_index];
-            for (int i = 0; i < swi.num_bits; i++) {
-                bool val = (pi.switch_mask & (1 << ((swi.num_bits - 1) - i))) != 0;
-                int8_t &cbit = config.at(swi.y).at(swi.x).at(swi.cbits[i].row).at(swi.cbits[i].col);
-                if (bool(cbit) != 0)
-                    NPNR_ASSERT(false);
-                cbit = val;
+            int sw_bel_idx = swi.bel;
+            if (sw_bel_idx >= 0) {
+                const BelInfoPOD &beli = ci.bel_data[sw_bel_idx];
+                const TileInfoPOD &ti = bi.tiles_nonrouting[TILE_LOGIC];
+                BelId sw_bel;
+                sw_bel.index = sw_bel_idx;
+                NPNR_ASSERT(ctx->getBelType(sw_bel) == TYPE_ICESTORM_LC);
+                BelPin input = get_one_bel_pin(ctx, ctx->getPipSrcWire(pip));
+                BelPin output = get_one_bel_pin(ctx, ctx->getPipDstWire(pip));
+                NPNR_ASSERT(input.bel == sw_bel);
+                NPNR_ASSERT(output.bel == sw_bel && output.pin == PIN_O);
+                unsigned lut_init;
+                switch (input.pin) {
+                case PIN_I0:
+                    lut_init = 2;
+                    break;
+                case PIN_I1:
+                    lut_init = 4;
+                    break;
+                case PIN_I2:
+                    lut_init = 16;
+                    break;
+                case PIN_I3:
+                    lut_init = 256;
+                    break;
+                default:
+                    NPNR_ASSERT_FALSE("bad feedthru LUT input");
+                }
+                std::vector<bool> lc(20, false);
+                for (int i = 0; i < 16; i++) {
+                    if ((lut_init >> i) & 0x1)
+                        lc.at(lut_perm.at(i)) = true;
+                }
+
+                for (int i = 0; i < 20; i++)
+                    set_config(ti, config.at(beli.y).at(beli.x), "LC_" + std::to_string(beli.z), lc.at(i), i);
+            } else {
+                for (int i = 0; i < swi.num_bits; i++) {
+                    bool val = (pi.switch_mask & (1 << ((swi.num_bits - 1) - i))) != 0;
+                    int8_t &cbit = config.at(swi.y).at(swi.x).at(swi.cbits[i].row).at(swi.cbits[i].col);
+                    if (bool(cbit) != 0)
+                        NPNR_ASSERT(false);
+                    cbit = val;
+                }
             }
         }
     }
@@ -295,9 +346,7 @@ void write_asc(const Context *ctx, std::ostream &out)
             bool carry_enable = get_param_or_def(cell.second.get(), ctx->id("CARRY_ENABLE"));
             std::vector<bool> lc(20, false);
             // From arachne-pnr
-            static std::vector<int> lut_perm = {
-                    4, 14, 15, 5, 6, 16, 17, 7, 3, 13, 12, 2, 1, 11, 10, 0,
-            };
+
             for (int i = 0; i < 16; i++) {
                 if ((lut_init >> i) & 0x1)
                     lc.at(lut_perm.at(i)) = true;
diff --git a/ice40/chipdb.py b/ice40/chipdb.py
index 3f9f4e65..a0d7f03c 100644
--- a/ice40/chipdb.py
+++ b/ice40/chipdb.py
@@ -271,6 +271,16 @@ def pipdelay(src_idx, dst_idx, db):
     if re.match(r"ram/(MASK|RADDR|WADDR|WDATA)_", dst[2]):
         return db["InMux.I.O"]
 
+    if re.match(r"lutff_\d+/out", dst[2]):
+        if re.match(r"lutff_\d+/in_0", src[2]):
+            return db["LogicCell40.in0.lcout"]
+        if re.match(r"lutff_\d+/in_1", src[2]):
+            return db["LogicCell40.in1.lcout"]
+        if re.match(r"lutff_\d+/in_2", src[2]):
+            return db["LogicCell40.in2.lcout"]
+        if re.match(r"lutff_\d+/in_3", src[2]):
+            return db["LogicCell40.in3.lcout"]
+
     print(src, dst, src_idx, dst_idx, src_type, dst_type, file=sys.stderr)
     assert 0
 
@@ -316,12 +326,12 @@ with open(args.filename, "r") as f:
 
         if line[0] == ".buffer":
             mode = ("buffer", int(line[3]), int(line[1]), int(line[2]))
-            switches.append((line[3], int(line[1]), int(line[2]), line[4:]))
+            switches.append((int(line[1]), int(line[2]), line[4:], -1))
             continue
 
         if line[0] == ".routing":
             mode = ("routing", int(line[3]), int(line[1]), int(line[2]))
-            switches.append((line[3], int(line[1]), int(line[2]), line[4:]))
+            switches.append((int(line[1]), int(line[2]), line[4:], -1))
             continue
 
         if line[0] == ".io_tile":
@@ -499,6 +509,22 @@ def add_wire(x, y, name):
     wire_names_r[wire_idx] = wname
     wire_segments[wire_idx] = dict()
 
+def add_switch(x, y, bel=-1):
+    switches.append((x, y, [], bel))
+
+def add_pip(src, dst):
+    x, y, _, _ = switches[-1]
+
+    if src not in wire_downhill:
+        wire_downhill[src] = set()
+    wire_downhill[src].add(dst)
+
+    if dst not in wire_uphill:
+        wire_uphill[dst] = set()
+    wire_uphill[dst].add(src)
+
+    pip_xy[(src, dst)] = (x, y, 0, len(switches) - 1)
+
 # Add virtual padin wires
 for i in range(8):
     add_wire(0, 0, "padin_%d" % i)
@@ -555,6 +581,13 @@ def add_bel_lc(x, y, z):
     if wire_lout is not None:
         add_bel_output(bel, wire_lout, "LO")
 
+    # route-through LUTs
+    add_switch(x, y, bel)
+    add_pip(wire_in_0, wire_out)
+    add_pip(wire_in_1, wire_out)
+    add_pip(wire_in_2, wire_out)
+    add_pip(wire_in_3, wire_out)
+
 def add_bel_io(x, y, z):
     bel = len(bel_name)
     bel_name.append("X%d/Y%d/io%d" % (x, y, z))
@@ -1050,7 +1083,7 @@ for info in pipinfo:
 
 switchinfo = []
 for switch in switches:
-    dst, x, y, bits = switch
+    x, y, bits, bel = switch
     bitlist = []
     for b in bits:
         m = cbit_re.match(b)
@@ -1060,11 +1093,13 @@ for switch in switches:
     si["x"] = x
     si["y"] = y
     si["bits"] = bitlist
+    si["bel"] = bel
     switchinfo.append(si)
 
 bba.l("switch_data_%s" % dev_name, "SwitchInfoPOD")
 for info in switchinfo:
     bba.u32(len(info["bits"]), "num_bits")
+    bba.u32(info["bel"], "bel")
     bba.u8(info["x"], "x")
     bba.u8(info["y"], "y")
     for i in range(5):
diff --git a/ice40/gfx.cc b/ice40/gfx.cc
index 0798862a..924fe964 100644
--- a/ice40/gfx.cc
+++ b/ice40/gfx.cc
@@ -696,11 +696,30 @@ void gfxTilePip(std::vector<GraphicElement> &g, int x, int y, GfxTileWireId src,
 {
     float x1, y1, x2, y2;
 
-    if (getWireXY_main(src, x1, y1) && getWireXY_main(dst, x2, y2))
+    if (getWireXY_main(src, x1, y1) && getWireXY_main(dst, x2, y2)) {
         pipGfx(g, x, y, x1, y1, x2, y2, main_swbox_x1, main_swbox_y1, main_swbox_x2, main_swbox_y2, style);
+        return;
+    }
 
-    if (getWireXY_local(src, x1, y1) && getWireXY_local(dst, x2, y2))
+    if (getWireXY_local(src, x1, y1) && getWireXY_local(dst, x2, y2)) {
         pipGfx(g, x, y, x1, y1, x2, y2, local_swbox_x1, local_swbox_y1, local_swbox_x2, local_swbox_y2, style);
+        return;
+    }
+
+    if (TILE_WIRE_LUTFF_0_IN_0 <= src && src <= TILE_WIRE_LUTFF_7_IN_3 && TILE_WIRE_LUTFF_0_OUT <= dst && dst <= TILE_WIRE_LUTFF_7_OUT) {
+        int lut_idx = (src - TILE_WIRE_LUTFF_0_IN_0) / 4;
+        int in_idx = (src - TILE_WIRE_LUTFF_0_IN_0) % 4;
+
+        GraphicElement el;
+        el.type = GraphicElement::TYPE_ARROW;
+        el.style = style;
+        el.x1 = x + logic_cell_x1;
+        el.x2 = x + logic_cell_x2;
+        el.y1 = y + (logic_cell_y1 + logic_cell_y2) / 2 - 0.0075 + (0.005 * in_idx) + lut_idx * logic_cell_pitch;
+        el.y2 = y + (logic_cell_y1 + logic_cell_y2) / 2 + lut_idx * logic_cell_pitch;
+        g.push_back(el);
+        return;
+    }
 
     if (src == TILE_WIRE_CARRY_IN && dst == TILE_WIRE_CARRY_IN_MUX) {
         GraphicElement el;
@@ -711,6 +730,7 @@ void gfxTilePip(std::vector<GraphicElement> &g, int x, int y, GfxTileWireId src,
         el.y1 = y + 0.01;
         el.y2 = y + 0.02;
         g.push_back(el);
+        return;
     }
 }