diff --git a/autotest.c b/autotest.c
index c27d916..a9a48a1 100644
--- a/autotest.c
+++ b/autotest.c
@@ -967,7 +967,7 @@ static int test_logic(struct test_state* tstate, int y, int x, int type_idx,
 {
 	struct fpga_device* dev;
 	net_idx_t pinw_nets[MAX_NUM_PINW];
-	int i, lut, rc;
+	int i, lut, latch_logic, rc;
 
 	if (tstate->dry_run) {
 		for (lut = LUT_A; lut <= LUT_D; lut++) {
@@ -987,6 +987,14 @@ static int test_logic(struct test_state* tstate, int y, int x, int type_idx,
 		fdev_print_required_pins(tstate->model, y, x,
 			DEV_LOGIC, type_idx);
 	}
+	latch_logic = 0;
+	for (lut = LUT_A; lut <= LUT_D; lut++) {
+		if (logic_cfg->a2d[lut].ff == FF_AND2L
+		    || logic_cfg->a2d[lut].ff == FF_OR2L) {
+			latch_logic = 1;
+			break;
+		}
+	}
 
 	// add stub nets for each required pin
 	dev = fdev_p(tstate->model, y, x, DEV_LOGIC, type_idx);
@@ -1009,7 +1017,10 @@ static int test_logic(struct test_state* tstate, int y, int x, int type_idx,
 			&& *dev->u.logic.a2d[LUT_C].lut5)
 		    || (dev->pinw_req_for_cfg[i] == LI_D6
 			&& dev->u.logic.a2d[LUT_D].lut5
-			&& *dev->u.logic.a2d[LUT_D].lut5)) {
+			&& *dev->u.logic.a2d[LUT_D].lut5)
+		    || (latch_logic
+			&& (dev->pinw_req_for_cfg[i] == LI_CLK
+			    || dev->pinw_req_for_cfg[i] == LI_CE))) {
 			rc = fnet_route_to_inpins(tstate->model, pinw_nets[i], "VCC_WIRE");
 			if (rc) FAIL(rc);
 		}
@@ -1234,16 +1245,18 @@ static int test_logic_config(struct test_state* tstate)
 					idx_enum[type_i], &logic_cfg);
 				if (rc) FAIL(rc);
 
-				// lut6, mux-out
-				memset(&logic_cfg, 0, sizeof(logic_cfg));
-				logic_cfg.a2d[lut].lut6 = "A1";
-				logic_cfg.a2d[lut].out_mux = MUX_O6;
-
-				rc = test_logic(tstate, y, x_enum[x_i],
-					idx_enum[type_i], &logic_cfg);
-				if (rc) FAIL(rc);
-
 				if (idx_enum[type_i] == DEV_LOG_M_OR_L) {
+					// lut6, mux-out
+					// O6 over mux-out seems not supported
+					// on an X device.
+					memset(&logic_cfg, 0, sizeof(logic_cfg));
+					logic_cfg.a2d[lut].lut6 = "A1";
+					logic_cfg.a2d[lut].out_mux = MUX_O6;
+
+					rc = test_logic(tstate, y, x_enum[x_i],
+						idx_enum[type_i], &logic_cfg);
+					if (rc) FAIL(rc);
+
 					// . out_mux=xor
 					logic_cfg.a2d[lut].out_mux = MUX_XOR;
 					if (lut == LUT_A) {
@@ -1317,6 +1330,64 @@ static int test_logic_config(struct test_state* tstate)
 					}
 				}
 
+				// lut6, latch-out
+				memset(&logic_cfg, 0, sizeof(logic_cfg));
+				logic_cfg.a2d[lut].lut6 = "A1";
+				logic_cfg.a2d[lut].ff = FF_LATCH;
+				logic_cfg.a2d[lut].ff_mux = MUX_O6;
+				logic_cfg.a2d[lut].ff_srinit = FF_SRINIT0;
+
+				logic_cfg.clk_inv = CLKINV_CLK;
+				logic_cfg.sync_attr = SYNCATTR_ASYNC;
+
+				rc = test_logic(tstate, y, x_enum[x_i],
+					idx_enum[type_i], &logic_cfg);
+				if (rc) FAIL(rc);
+
+				//
+				// AND and OR latches are physically a normal
+				// latch, but with additional configuration
+				// constraints:
+				//
+				// 1. ce and clk must be driven high/vcc
+				// 2. the clk must be inverted before the latch (clk_b)
+				// 3. srinit must be 0 for AND, 1 for OR
+				//
+
+				// lut6, and-latch
+				memset(&logic_cfg, 0, sizeof(logic_cfg));
+				logic_cfg.a2d[lut].lut6 = "A1";
+				logic_cfg.a2d[lut].ff = FF_AND2L;
+				logic_cfg.a2d[lut].ff_mux = MUX_O6;
+				// AND2L requires SRINIT=0
+				logic_cfg.a2d[lut].ff_srinit = FF_SRINIT0;
+
+				logic_cfg.clk_inv = CLKINV_B;
+				logic_cfg.sync_attr = SYNCATTR_ASYNC;
+				logic_cfg.ce_used = 1;
+				logic_cfg.sr_used = 1;
+
+				rc = test_logic(tstate, y, x_enum[x_i],
+					idx_enum[type_i], &logic_cfg);
+				if (rc) FAIL(rc);
+
+				// lut6, or-latch
+				memset(&logic_cfg, 0, sizeof(logic_cfg));
+				logic_cfg.a2d[lut].lut6 = "A1";
+				logic_cfg.a2d[lut].ff = FF_OR2L;
+				logic_cfg.a2d[lut].ff_mux = MUX_O6;
+				// OR2L requires SRINIT=1
+				logic_cfg.a2d[lut].ff_srinit = FF_SRINIT1;
+
+				logic_cfg.clk_inv = CLKINV_B;
+				logic_cfg.sync_attr = SYNCATTR_ASYNC;
+				logic_cfg.ce_used = 1;
+				logic_cfg.sr_used = 1;
+
+				rc = test_logic(tstate, y, x_enum[x_i],
+					idx_enum[type_i], &logic_cfg);
+				if (rc) FAIL(rc);
+
 				// x, ff-out
 				memset(&logic_cfg, 0, sizeof(logic_cfg));
 				logic_cfg.a2d[lut].ff = FF_FF;
@@ -1337,12 +1408,14 @@ static int test_logic_config(struct test_state* tstate)
 				if (rc) FAIL(rc);
 				logic_cfg.a2d[lut].out_used = 0;
 
-				// . o6-outmux
-				logic_cfg.a2d[lut].out_mux = MUX_O6;
-				rc = test_logic(tstate, y, x_enum[x_i],
-					idx_enum[type_i], &logic_cfg);
-				if (rc) FAIL(rc);
-				logic_cfg.a2d[lut].out_mux = 0;
+				if (idx_enum[type_i] == DEV_LOG_M_OR_L) {
+					// . o6-outmux
+					logic_cfg.a2d[lut].out_mux = MUX_O6;
+					rc = test_logic(tstate, y, x_enum[x_i],
+						idx_enum[type_i], &logic_cfg);
+					if (rc) FAIL(rc);
+					logic_cfg.a2d[lut].out_mux = 0;
+				}
 
 				//
 				// lut5/6 pairs
@@ -1388,17 +1461,17 @@ static int test_logic_config(struct test_state* tstate)
 					x_enum[x_i], idx_enum[type_i], &logic_cfg);
 				if (rc) FAIL(rc);
 
-				// . change from out_mux/5q to ff_mux
-				logic_cfg.a2d[lut].out_mux = 0;
-				logic_cfg.a2d[lut].ff5_srinit = 0;
-				logic_cfg.a2d[lut].ff = FF_FF;
-				logic_cfg.a2d[lut].ff_mux = MUX_O5;
-				logic_cfg.a2d[lut].ff_srinit = FF_SRINIT0;
-				rc = test_logic(tstate, y, x_enum[x_i],
-					idx_enum[type_i], &logic_cfg);
-				if (rc) FAIL(rc);
-
 				if (idx_enum[type_i] == DEV_LOG_M_OR_L) {
+					// . change from out_mux/5q to ff_mux
+					logic_cfg.a2d[lut].out_mux = 0;
+					logic_cfg.a2d[lut].ff5_srinit = 0;
+					logic_cfg.a2d[lut].ff = FF_FF;
+					logic_cfg.a2d[lut].ff_mux = MUX_O5;
+					logic_cfg.a2d[lut].ff_srinit = FF_SRINIT0;
+					rc = test_logic(tstate, y, x_enum[x_i],
+						idx_enum[type_i], &logic_cfg);
+					if (rc) FAIL(rc);
+
 					// . add out_mux=cy cy0=x
 					logic_cfg.a2d[lut].out_mux = MUX_CY;
 					logic_cfg.a2d[lut].cy0 = CY0_X;
@@ -1422,7 +1495,6 @@ static int test_logic_config(struct test_state* tstate)
 						if (rc) FAIL(rc);
 					}
 				}
-//break;
 			}
 			if (idx_enum[type_i] != DEV_LOG_M_OR_L)
 				continue;
@@ -1524,7 +1596,6 @@ static int test_logic_config(struct test_state* tstate)
 			rc = test_logic(tstate, y, x_enum[x_i],
 				idx_enum[type_i], &logic_cfg);
 			if (rc) FAIL(rc);
-//goto out;
 		}
 	}
 out:
diff --git a/blinking_led.c b/blinking_led.c
index 5e0da19..44dba8d 100644
--- a/blinking_led.c
+++ b/blinking_led.c
@@ -9,6 +9,22 @@
 #include "floorplan.h"
 #include "control.h"
 
+/*
+   This C design corresponds to the following Verilog:
+  
+   module blinking(input clk, output led);
+
+   // synthesis attribute LOC clk "P55 | IOSTANDARD = LVCMOS33"
+   // synthesis attribute LOC led "P48 | SLEW = QUIETIO | DRIVE = 8"
+
+     reg [14:0] counter;
+     always @(posedge clk) counter <= counter + 1;
+     assign led = counter[14];
+
+   endmodule
+
+*/
+
 int main(int argc, char** argv)
 {
 	struct fpga_model model;
diff --git a/hello_world.c b/hello_world.c
index 5e0da19..db0de19 100644
--- a/hello_world.c
+++ b/hello_world.c
@@ -9,6 +9,20 @@
 #include "floorplan.h"
 #include "control.h"
 
+/*
+   This C design corresponds to the following Verilog:
+  
+   module ver_and(input a, b, output y);
+
+   // synthesis attribute LOC a "P45"
+   // synthesis attribute LOC b "P46"
+   // synthesis attribute LOC y "P48 | IOSTANDARD = LVCMOS33 | SLEW = SLOW | DRIVE = 12"
+
+    assign y = a & b;
+
+   endmodule
+*/
+
 int main(int argc, char** argv)
 {
 	struct fpga_model model;
diff --git a/libs/bit_regs.c b/libs/bit_regs.c
index 0242505..e291e02 100644
--- a/libs/bit_regs.c
+++ b/libs/bit_regs.c
@@ -762,7 +762,7 @@ static void printf_v64_mi20(const uint8_t* bits, int row, int major)
 				if (u64 & (1ULL << i))
 					num_bits_on++;
 			}
-			if (num_bits_on < 3) {
+			if (num_bits_on < 5) {
 				for (i = 0; i < 64; i++) {
 					if (!(u64 & (1ULL << i)))
 						continue;
diff --git a/libs/control.c b/libs/control.c
index 0658b5e..ddf2731 100644
--- a/libs/control.c
+++ b/libs/control.c
@@ -458,15 +458,14 @@ int fdev_logic_setconf(struct fpga_model* model, int y, int x,
 				lut, 5, logic_cfg->a2d[lut].lut5, ZTERM);
 			if (rc) FAIL(rc);
 		}
-		if (logic_cfg->a2d[lut].ff == FF_FF) {
+		if (logic_cfg->a2d[lut].ff) {
 			if (!logic_cfg->a2d[lut].ff_mux
 			    || !logic_cfg->a2d[lut].ff_srinit)
 				FAIL(EINVAL);
-			dev->u.logic.a2d[lut].ff = FF_FF;
+			dev->u.logic.a2d[lut].ff = logic_cfg->a2d[lut].ff;
 			dev->u.logic.a2d[lut].ff_mux = logic_cfg->a2d[lut].ff_mux;
 			dev->u.logic.a2d[lut].ff_srinit = logic_cfg->a2d[lut].ff_srinit;
-		} else if (logic_cfg->a2d[lut].ff)
-			FAIL(EINVAL);
+		}
 		if (logic_cfg->a2d[lut].out_mux)
 			dev->u.logic.a2d[lut].out_mux = logic_cfg->a2d[lut].out_mux;
 		if (logic_cfg->a2d[lut].ff5_srinit)
@@ -841,6 +840,18 @@ int fdev_set_required_pins(struct fpga_model* model, int y, int x, int type,
 		}
 		if (dev->u.logic.precyinit == PRECYINIT_AX)
 			add_req_inpin(dev, LI_AX);
+		if (dev->u.logic.a2d[LUT_A].out_mux == MUX_F7
+		    || dev->u.logic.a2d[LUT_A].ff_mux == MUX_F7)
+			add_req_inpin(dev, LI_AX);
+		if (dev->u.logic.a2d[LUT_C].out_mux == MUX_F7
+		    || dev->u.logic.a2d[LUT_C].ff_mux == MUX_F7)
+			add_req_inpin(dev, LI_CX);
+		if (dev->u.logic.a2d[LUT_B].out_mux == MUX_F8
+		    || dev->u.logic.a2d[LUT_B].ff_mux == MUX_F8) {
+			add_req_inpin(dev, LI_AX);
+			add_req_inpin(dev, LI_BX);
+			add_req_inpin(dev, LI_CX);
+		}
 		for (i = LUT_A; i <= LUT_D; i++) {
 			if (dev->u.logic.a2d[i].out_used) {
 				// LO_A..LO_D are in sequence
@@ -1575,16 +1586,24 @@ int fpga_switch_chain(struct sw_chain* ch)
 					= ch->set.sw[ch->set.len-1];
 			}
 
-			idx = fpga_switch_next(ch->model, ch->y, ch->x,
+			ch->set.sw[ch->set.len-1] = fpga_switch_next(
+				ch->model, ch->y, ch->x,
 				ch->set.sw[ch->set.len-1], ch->from_to);
-			if (idx != NO_SWITCH) {
-				if (fpga_switch_is_used(ch->model, ch->y, ch->x, idx))
+			if (ch->set.sw[ch->set.len-1] != NO_SWITCH) {
+				if (fpga_switch_is_used(ch->model, ch->y,
+					ch->x, ch->set.sw[ch->set.len-1])) {
+#ifdef DBG_ENUM_SWITCH
+					printf(" skipping used %s\n",
+					  fpga_switch_print(ch->model, ch->y,
+					  ch->x, ch->set.sw[ch->set.len-1]));
+#endif
 					continue;
+				}
 #ifdef DBG_ENUM_SWITCH
 				printf(" found %s\n", fpga_switch_print(
-					ch->model, ch->y, ch->x, idx));
+				  ch->model, ch->y, ch->x,
+				  ch->set.sw[ch->set.len-1]));
 #endif
-				ch->set.sw[ch->set.len-1] = idx;
 				break;
 			}
 
@@ -2238,11 +2257,14 @@ int fnet_route_to_inpins(struct fpga_model* model, net_idx_t net_i,
 		if ((net_p->el[i].idx & NET_IDX_MASK) >= dev_p->num_pinw_in)
 			// skip outpin
 			continue;
+
 		rc = froute_direct(model, net_p->el[i].y, net_p->el[i].x-1,
 			from_i, net_p->el[i].y, net_p->el[i].x,
 			dev_p->pinw[net_p->el[i].idx & NET_IDX_MASK],
 			&start_set, &end_set);
 		if (rc) FAIL(rc);
+		if (!start_set.len || !end_set.len)
+			HERE();
 		rc = fnet_add_sw(model, net_i, net_p->el[i].y,
 			net_p->el[i].x-1, start_set.sw, start_set.len);
 		if (rc) FAIL(rc);
@@ -2268,7 +2290,7 @@ int froute_direct(struct fpga_model* model, int start_y, int start_x,
 	if (!end_switches.len) FAIL(EINVAL);
 
 	rc = construct_sw_conns(&conns, model, start_y, start_x, start_pt,
-		SW_FROM, /*max_depth*/ 1);
+		SW_FROM, /*max_depth*/ 2);
 	if (rc) FAIL(rc);
 
 	while (fpga_switch_conns(&conns) != NO_CONN) {