diff --git a/Makefile b/Makefile
index e56db424d..692b19826 100644
--- a/Makefile
+++ b/Makefile
@@ -115,7 +115,7 @@ LDFLAGS += -rdynamic
 LDLIBS += -lrt
 endif
 
-YOSYS_VER := 0.9+1
+YOSYS_VER := 0.9+36
 GIT_REV := $(shell cd $(YOSYS_SRC) && git rev-parse --short HEAD 2> /dev/null || echo UNKNOWN)
 OBJS = kernel/version_$(GIT_REV).o
 
diff --git a/passes/sat/async2sync.cc b/passes/sat/async2sync.cc
index d045d0dcb..24ae6e448 100644
--- a/passes/sat/async2sync.cc
+++ b/passes/sat/async2sync.cc
@@ -39,7 +39,7 @@ struct Async2syncPass : public Pass {
 		log("reset value in the next cycle regardless of the data-in value at the time of\n");
 		log("the clock edge.\n");
 		log("\n");
-		log("Currently only $adff and $dffsr cells are supported by this pass.\n");
+		log("Currently only $adff, $dffsr, and $dlatch cells are supported by this pass.\n");
 		log("\n");
 	}
 	void execute(std::vector<std::string> args, RTLIL::Design *design) YS_OVERRIDE
@@ -169,6 +169,41 @@ struct Async2syncPass : public Pass {
 					cell->type = "$dff";
 					continue;
 				}
+
+				if (cell->type.in("$dlatch"))
+				{
+					bool en_pol = cell->parameters["\\EN_POLARITY"].as_bool();
+
+					SigSpec sig_en = cell->getPort("\\EN");
+					SigSpec sig_d = cell->getPort("\\D");
+					SigSpec sig_q = cell->getPort("\\Q");
+
+					log("Replacing %s.%s (%s): EN=%s, D=%s, Q=%s\n",
+							log_id(module), log_id(cell), log_id(cell->type),
+							log_signal(sig_en), log_signal(sig_d), log_signal(sig_q));
+
+					Const init_val;
+					for (int i = 0; i < GetSize(sig_q); i++) {
+						SigBit bit = sigmap(sig_q[i]);
+						init_val.bits.push_back(initbits.count(bit) ? initbits.at(bit) : State::Sx);
+						del_initbits.insert(bit);
+					}
+
+					Wire *new_q = module->addWire(NEW_ID, GetSize(sig_q));
+					new_q->attributes["\\init"] = init_val;
+
+					if (en_pol) {
+						module->addMux(NEW_ID, new_q, sig_d, sig_en, sig_q);
+					} else {
+						module->addMux(NEW_ID, sig_d, new_q, sig_en, sig_q);
+					}
+
+					cell->setPort("\\Q", new_q);
+					cell->unsetPort("\\EN");
+					cell->unsetParam("\\EN_POLARITY");
+					cell->type = "$ff";
+					continue;
+				}
 			}
 
 			for (auto wire : module->wires())
diff --git a/passes/techmap/shregmap.cc b/passes/techmap/shregmap.cc
index 9da69e8ba..be00e5030 100644
--- a/passes/techmap/shregmap.cc
+++ b/passes/techmap/shregmap.cc
@@ -191,7 +191,7 @@ struct ShregmapWorker
 				IdString q_port = opts.ffcells.at(c1->type).second;
 
 				auto c1_conn = c1->connections();
-				auto c2_conn = c1->connections();
+				auto c2_conn = c2->connections();
 
 				c1_conn.erase(d_port);
 				c1_conn.erase(q_port);
diff --git a/techlibs/ecp5/Makefile.inc b/techlibs/ecp5/Makefile.inc
index c41d16076..2143acae6 100644
--- a/techlibs/ecp5/Makefile.inc
+++ b/techlibs/ecp5/Makefile.inc
@@ -2,6 +2,8 @@
 OBJS += techlibs/ecp5/synth_ecp5.o techlibs/ecp5/ecp5_ffinit.o \
         techlibs/ecp5/ecp5_gsr.o
 
+$(eval $(call add_share_file,share/ecp5,techlibs/ecp5/cells_ff.vh))
+$(eval $(call add_share_file,share/ecp5,techlibs/ecp5/cells_io.vh))
 $(eval $(call add_share_file,share/ecp5,techlibs/ecp5/cells_map.v))
 $(eval $(call add_share_file,share/ecp5,techlibs/ecp5/cells_sim.v))
 $(eval $(call add_share_file,share/ecp5,techlibs/ecp5/cells_bb.v))
diff --git a/techlibs/ecp5/cells_ff.vh b/techlibs/ecp5/cells_ff.vh
new file mode 100644
index 000000000..0c9689ebd
--- /dev/null
+++ b/techlibs/ecp5/cells_ff.vh
@@ -0,0 +1,40 @@
+// Diamond flip-flops
+module FD1P3AX(input     D, SP, CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(0),  .CE(SP), .DI(D), .Q(Q)); endmodule
+module FD1P3AY(input     D, SP, CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(0),  .CE(SP), .DI(D), .Q(Q)); endmodule
+module FD1P3BX(input PD, D, SP, CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module FD1P3DX(input CD, D, SP, CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module FD1P3IX(input CD, D, SP, CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module FD1P3JX(input PD, D, SP, CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module FD1S3AX(input     D,     CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("1"),  .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(0),           .DI(D), .Q(Q)); endmodule
+module FD1S3AY(input     D,     CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("1"),  .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(0),           .DI(D), .Q(Q)); endmodule
+module FD1S3BX(input PD, D,     CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("1"),  .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(PD),          .DI(D), .Q(Q)); endmodule
+module FD1S3DX(input CD, D,     CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("1"),  .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))        _TECHMAP_REPLACE_ (.CLK(CK), .LSR(CD),          .DI(D), .Q(Q)); endmodule
+module FD1S3IX(input CD, D,     CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("1"),  .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(CD),          .DI(D), .Q(Q)); endmodule
+module FD1S3JX(input PD, D,     CK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("1"),  .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(PD),          .DI(D), .Q(Q)); endmodule
+
+// TODO: Diamond latches
+// module FL1P3AY(); endmodule
+// module FL1P3AZ(); endmodule
+// module FL1P3BX(); endmodule
+// module FL1P3DX(); endmodule
+// module FL1P3IY(); endmodule
+// module FL1P3JY(); endmodule
+// module FL1S3AX(); endmodule
+// module FL1S3AY(); endmodule
+
+// Diamond I/O registers
+module IFS1P3BX(input PD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("ASYNC"))       _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module IFS1P3DX(input CD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))       _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module IFS1P3IX(input CD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE")) _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module IFS1P3JX(input PD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("LSR_OVER_CE")) _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
+
+module OFS1P3BX(input PD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("ASYNC"))       _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module OFS1P3DX(input CD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))       _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module OFS1P3IX(input CD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE")) _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
+module OFS1P3JX(input PD, D, SP, SCLK, output Q); parameter GSR = "ENABLED"; TRELLIS_FF #(.GSR(GSR), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"),   .SRMODE("LSR_OVER_CE")) _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
+
+// TODO: Diamond I/O latches
+// module IFS1S1B(input PD, D, SCLK, output Q); endmodule
+// module IFS1S1D(input CD, D, SCLK, output Q); endmodule
+// module IFS1S1I(input PD, D, SCLK, output Q); endmodule
+// module IFS1S1J(input CD, D, SCLK, output Q); endmodule
diff --git a/techlibs/ecp5/cells_io.vh b/techlibs/ecp5/cells_io.vh
new file mode 100644
index 000000000..02e66e8a5
--- /dev/null
+++ b/techlibs/ecp5/cells_io.vh
@@ -0,0 +1,14 @@
+// Diamond I/O buffers
+module IB   (input I,     output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("INPUT"))  _TECHMAP_REPLACE_ (.B(I), .O(O)); endmodule
+module IBPU (input I,     output O); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("INPUT"))  _TECHMAP_REPLACE_ (.B(I), .O(O)); endmodule
+module IBPD (input I,     output O); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("INPUT"))  _TECHMAP_REPLACE_ (.B(I), .O(O)); endmodule
+module OB   (input I,     output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(O), .I(I)); endmodule
+module OBZ  (input I, T,  output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(O), .I(I), .T(T)); endmodule
+module OBZPU(input I, T,  output O); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(O), .I(I), .T(T)); endmodule
+module OBZPD(input I, T,  output O); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(O), .I(I), .T(T)); endmodule
+module OBCO (input I,     output OT, OC); OLVDS olvds (.A(I), .Z(OT), .ZN(OC)); endmodule
+module BB   (input I, T,  output O, inout B); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("BIDIR")) _TECHMAP_REPLACE_ (.B(B), .I(I), .O(O), .T(T)); endmodule
+module BBPU (input I, T,  output O, inout B); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("BIDIR")) _TECHMAP_REPLACE_ (.B(B), .I(I), .O(O), .T(T)); endmodule
+module BBPD (input I, T,  output O, inout B); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("BIDIR")) _TECHMAP_REPLACE_ (.B(B), .I(I), .O(O), .T(T)); endmodule
+module ILVDS(input A, AN, output Z    ); TRELLIS_IO #(.DIR("INPUT"))  _TECHMAP_REPLACE_ (.B(A), .O(Z)); endmodule
+module OLVDS(input A,     output Z, ZN); TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(Z), .I(A)); endmodule
diff --git a/techlibs/ecp5/cells_map.v b/techlibs/ecp5/cells_map.v
index 0a92d906d..71ae9237b 100644
--- a/techlibs/ecp5/cells_map.v
+++ b/techlibs/ecp5/cells_map.v
@@ -47,59 +47,8 @@ module  \$__DFFSE_NP1 (input D, C, E, R, output Q); TRELLIS_FF #(.GSR("AUTO"), .
 module  \$__DFFSE_PP0 (input D, C, E, R, output Q); TRELLIS_FF #(.GSR("AUTO"), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(C), .CE(E), .LSR(R), .DI(D), .Q(Q)); endmodule
 module  \$__DFFSE_PP1 (input D, C, E, R, output Q); TRELLIS_FF #(.GSR("AUTO"), .CEMUX("CE"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(C), .CE(E), .LSR(R), .DI(D), .Q(Q)); endmodule
 
-// TODO: Diamond flip-flops
-// module FD1P3AX(); endmodule
-// module FD1P3AY(); endmodule
-// module FD1P3BX(); endmodule
-// module FD1P3DX(); endmodule
-// module FD1P3IX(); endmodule
-// module FD1P3JX(); endmodule
-// module FD1S3AX(); endmodule
-// module FD1S3AY(); endmodule
-module FD1S3BX(input PD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("ASYNC"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(PD), .DI(D), .Q(Q)); endmodule
-module FD1S3DX(input CD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(CD), .DI(D), .Q(Q)); endmodule
-module FD1S3IX(input CD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(CD), .DI(D), .Q(Q)); endmodule
-module FD1S3JX(input PD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(CK), .LSR(PD), .DI(D), .Q(Q)); endmodule
-// module FL1P3AY(); endmodule
-// module FL1P3AZ(); endmodule
-// module FL1P3BX(); endmodule
-// module FL1P3DX(); endmodule
-// module FL1P3IY(); endmodule
-// module FL1P3JY(); endmodule
-// module FL1S3AX(); endmodule
-// module FL1S3AY(); endmodule
-
-// Diamond I/O buffers
-module IB   (input I, output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("INPUT")) _TECHMAP_REPLACE_ (.B(I), .O(O)); endmodule
-module IBPU (input I, output O); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("INPUT"))   _TECHMAP_REPLACE_ (.B(I), .O(O)); endmodule
-module IBPD (input I, output O); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("INPUT")) _TECHMAP_REPLACE_ (.B(I), .O(O)); endmodule
-module OB   (input I, output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(O), .I(I)); endmodule
-module OBZ  (input I, T, output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(O), .I(I), .T(T)); endmodule
-module OBZPU(input I, T, output O); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("OUTPUT"))   _TECHMAP_REPLACE_ (.B(O), .I(I), .T(T)); endmodule
-module OBZPD(input I, T, output O); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(O), .I(I), .T(T)); endmodule
-module OBCO (input I, output OT, OC); OLVDS _TECHMAP_REPLACE_ (.A(I), .Z(OT), .ZN(OC)); endmodule
-module BB   (input I, T, output O, inout B); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("BIDIR")) _TECHMAP_REPLACE_ (.B(B), .I(I), .O(O), .T(T)); endmodule
-module BBPU (input I, T, output O, inout B); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("BIDIR"))   _TECHMAP_REPLACE_ (.B(B), .I(I), .O(O), .T(T)); endmodule
-module BBPD (input I, T, output O, inout B); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("BIDIR")) _TECHMAP_REPLACE_ (.B(B), .I(I), .O(O), .T(T)); endmodule
-module ILVDS(input A, AN, output Z); TRELLIS_IO #(.DIR("INPUT"))  _TECHMAP_REPLACE_ (.B(A), .O(Z)); endmodule
-module OLVDS(input A, output Z, ZN); TRELLIS_IO #(.DIR("OUTPUT")) _TECHMAP_REPLACE_ (.B(Z), .I(A)); endmodule
-
-// Diamond I/O registers
-module IFS1P3BX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("ASYNC"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module IFS1P3DX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module IFS1P3IX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module IFS1P3JX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-
-module OFS1P3BX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("ASYNC"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module OFS1P3DX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module OFS1P3IX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module OFS1P3JX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("LSR_OVER_CE"))  _TECHMAP_REPLACE_ (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-
-// TODO: Diamond I/O latches
-// module IFS1S1B(input PD, D, SCLK, output Q); endmodule
-// module IFS1S1D(input CD, D, SCLK, output Q); endmodule
-// module IFS1S1I(input PD, D, SCLK, output Q); endmodule
-// module IFS1S1J(input CD, D, SCLK, output Q); endmodule
+`include "cells_ff.vh"
+`include "cells_io.vh"
 
 `ifndef NO_LUT
 module \$lut (A, Y);
diff --git a/techlibs/ecp5/cells_sim.v b/techlibs/ecp5/cells_sim.v
index dc8334acb..75a1aad1f 100644
--- a/techlibs/ecp5/cells_sim.v
+++ b/techlibs/ecp5/cells_sim.v
@@ -693,56 +693,9 @@ module DP16KD(
 	parameter INITVAL_3F = 320'h00000000000000000000000000000000000000000000000000000000000000000000000000000000;
 endmodule
 
-// TODO: Diamond flip-flops
-// module FD1P3AX(); endmodule
-// module FD1P3AY(); endmodule
-// module FD1P3BX(); endmodule
-// module FD1P3DX(); endmodule
-// module FD1P3IX(); endmodule
-// module FD1P3JX(); endmodule
-// module FD1S3AX(); endmodule
-// module FD1S3AY(); endmodule
-module FD1S3BX(input PD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("ASYNC"))  tff (.CLK(CK), .LSR(PD), .DI(D), .Q(Q)); endmodule
-module FD1S3DX(input CD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))  tff (.CLK(CK), .LSR(CD), .DI(D), .Q(Q)); endmodule
-module FD1S3IX(input CD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  tff (.CLK(CK), .LSR(CD), .DI(D), .Q(Q)); endmodule
-module FD1S3JX(input PD, D, CK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("LSR_OVER_CE"))  tff (.CLK(CK), .LSR(PD), .DI(D), .Q(Q)); endmodule
-// module FL1P3AY(); endmodule
-// module FL1P3AZ(); endmodule
-// module FL1P3BX(); endmodule
-// module FL1P3DX(); endmodule
-// module FL1P3IY(); endmodule
-// module FL1P3JY(); endmodule
-// module FL1S3AX(); endmodule
-// module FL1S3AY(); endmodule
+`ifndef NO_INCLUDES
 
-// Diamond I/O buffers
-module IB   (input I, output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("INPUT")) tio (.B(I), .O(O)); endmodule
-module IBPU (input I, output O); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("INPUT"))   tio (.B(I), .O(O)); endmodule
-module IBPD (input I, output O); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("INPUT")) tio (.B(I), .O(O)); endmodule
-module OB   (input I, output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("OUTPUT")) tio (.B(O), .I(I)); endmodule
-module OBZ  (input I, T, output O); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("OUTPUT")) tio (.B(O), .I(I), .T(T)); endmodule
-module OBZPU(input I, T, output O); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("OUTPUT"))   tio (.B(O), .I(I), .T(T)); endmodule
-module OBZPD(input I, T, output O); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("OUTPUT")) tio (.B(O), .I(I), .T(T)); endmodule
-module OBCO (input I, output OT, OC); OLVDS olvds (.A(I), .Z(OT), .ZN(OC)); endmodule
-module BB   (input I, T, output O, inout B); (* PULLMODE="NONE" *) TRELLIS_IO #(.DIR("BIDIR")) tio (.B(B), .I(I), .O(O), .T(T)); endmodule
-module BBPU (input I, T, output O, inout B); (* PULLMODE="UP"   *) TRELLIS_IO #(.DIR("BIDIR"))   tio (.B(B), .I(I), .O(O), .T(T)); endmodule
-module BBPD (input I, T, output O, inout B); (* PULLMODE="DOWN" *) TRELLIS_IO #(.DIR("BIDIR")) tio (.B(B), .I(I), .O(O), .T(T)); endmodule
-module ILVDS(input A, AN, output Z); TRELLIS_IO #(.DIR("INPUT"))  tio (.B(A), .O(Z)); endmodule
-module OLVDS(input A, output Z, ZN); TRELLIS_IO #(.DIR("OUTPUT")) tio (.B(Z), .I(A)); endmodule
+`include "cells_ff.vh"
+`include "cells_io.vh"
 
-// Diamond I/O registers
-module IFS1P3BX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("ASYNC"))  tff (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module IFS1P3DX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))  tff (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module IFS1P3IX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  tff (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module IFS1P3JX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("LSR_OVER_CE"))  tff (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-
-module OFS1P3BX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("ASYNC"))  tff (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module OFS1P3DX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("ASYNC"))  tff (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module OFS1P3IX(input CD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("RESET"), .SRMODE("LSR_OVER_CE"))  tff (.CLK(SCLK), .LSR(CD), .CE(SP), .DI(D), .Q(Q)); endmodule
-module OFS1P3JX(input PD, D, SP, SCLK, output Q); TRELLIS_FF #(.GSR("DISABLED"), .CEMUX("1"), .CLKMUX("CLK"), .LSRMUX("LSR"), .REGSET("SET"), .SRMODE("LSR_OVER_CE"))  tff (.CLK(SCLK), .LSR(PD), .CE(SP), .DI(D), .Q(Q)); endmodule
-
-// TODO: Diamond I/O latches
-// module IFS1S1B(input PD, D, SCLK, output Q); endmodule
-// module IFS1S1D(input CD, D, SCLK, output Q); endmodule
-// module IFS1S1I(input PD, D, SCLK, output Q); endmodule
-// module IFS1S1J(input CD, D, SCLK, output Q); endmodule
+`endif
diff --git a/techlibs/ecp5/tests/.gitignore b/techlibs/ecp5/tests/.gitignore
new file mode 100644
index 000000000..0e18132cc
--- /dev/null
+++ b/techlibs/ecp5/tests/.gitignore
@@ -0,0 +1 @@
+work_*
diff --git a/techlibs/ecp5/tests/test_diamond_ffs.py b/techlibs/ecp5/tests/test_diamond_ffs.py
new file mode 100644
index 000000000..1ed85ce8b
--- /dev/null
+++ b/techlibs/ecp5/tests/test_diamond_ffs.py
@@ -0,0 +1,82 @@
+import os
+import subprocess
+
+if not os.path.exists("work_ff"):
+	os.mkdir("work_ff")
+
+modules = []
+
+with open("../cells_ff.vh", "r") as f:
+	with open("work_ff/cells_ff_gate.v", "w") as g:
+		for line in f:
+			if not line.startswith("module"):
+				g.write(line)
+				continue
+			else:
+				spidx = line.find(" ")
+				bridx = line.find("(")
+				modname = line[spidx+1 : bridx]
+				g.write("module %s_gate" % modname)
+				g.write(line[bridx:])
+				inpidx = line.find("input ")
+				outpidx = line.find(", output")
+				modules.append((modname, [x.strip() for x in line[inpidx+6:outpidx].split(",")]))
+
+with open("work_ff/testbench.v", "w") as f:
+	print("""
+`timescale 1ns/ 1ps
+
+module testbench;
+reg pur = 0, clk, rst, cen, d;
+
+// Needed for Diamond sim models
+GSR GSR_INST (.GSR(1'b1));
+PUR PUR_INST (.PUR(pur));
+
+
+initial begin
+	$dumpfile("work_ff/ffs.vcd");
+	$dumpvars(0, testbench);
+	#5;
+	pur = 1;
+	#95;
+	repeat (2500) begin
+		{clk, rst, cen, d} = $random;
+		#10;
+		check_outputs;
+		#1;
+	end
+	$finish;
+end
+	""", file=f)
+
+	for modname, inputs in modules:
+		print("    wire %s_gold_q, %s_gate_q;"  % (modname, modname), file=f)
+		portconns = []
+		for inp in inputs:
+			if inp in ("SCLK", "CK"):
+				portconns.append(".%s(clk)" % inp)
+			elif inp in ("CD", "PD"):
+				portconns.append(".%s(rst)" % inp)
+			elif inp == "SP":
+				portconns.append(".%s(cen)" % inp)
+			elif inp == "D":
+				portconns.append(".%s(d)" % inp)
+			else:
+				assert False
+		portconns.append(".Q(%s_gold_q)" % modname)
+		print("    %s %s_gold_i (%s);" % (modname, modname, ", ".join(portconns)), file=f)
+		portconns[-1] = (".Q(%s_gate_q)" % modname)
+		print("    %s_gate %s_gate_i (%s);" % (modname, modname, ", ".join(portconns)), file=f)
+		print("", file=f)
+	print("    task check_outputs;", file=f)
+	print("        begin", file=f)
+	print("             if (%s_gold_q != %s_gate_q) $display(\"MISMATCH at %%1t:  %s_gold_q=%%b, %s_gate_q=%%b\", $time, %s_gold_q, %s_gate_q);" %
+			(modname, modname, modname, modname, modname, modname), file=f)
+	print("        end", file=f)
+	print("    endtask", file=f)
+	print("endmodule", file=f)
+
+diamond_models = "/usr/local/diamond/3.10_x64/cae_library/simulation/verilog/ecp5u"
+subprocess.call(["iverilog", "-s", "testbench", "-o", "work_ff/testbench", "-Dmixed_hdl", "-DNO_INCLUDES", "-y", diamond_models, "work_ff/cells_ff_gate.v", "../cells_sim.v", "work_ff/testbench.v"])
+subprocess.call(["vvp", "work_ff/testbench"])
diff --git a/techlibs/ice40/abc_hx.box b/techlibs/ice40/abc_hx.box
index c0ea742e2..3ea70bc91 100644
--- a/techlibs/ice40/abc_hx.box
+++ b/techlibs/ice40/abc_hx.box
@@ -3,11 +3,11 @@
 # NB: Inputs/Outputs must be ordered alphabetically
 #     (with exceptions for carry in/out)
 
-# Inputs: A B CI
+# Inputs: A B I0 I3 CI
 # Outputs: O CO
 #   (NB: carry chain input/output must be last
 #        input/output and have been moved there
 #        overriding the alphabetical ordering)
-$__ICE40_FULL_ADDER 1 1 3 2
-400 379 316
-259 231 126
+$__ICE40_CARRY_WRAPPER 1 1 5 2
+400 379 449 316 316
+259 231 -   -   126
diff --git a/techlibs/ice40/abc_lp.box b/techlibs/ice40/abc_lp.box
index d73b6d649..473e92fe9 100644
--- a/techlibs/ice40/abc_lp.box
+++ b/techlibs/ice40/abc_lp.box
@@ -3,11 +3,11 @@
 # NB: Inputs/Outputs must be ordered alphabetically
 #     (with exceptions for carry in/out)
 
-# Inputs: A B CI
+# Inputs: A B I0 I3 CI
 # Outputs: O CO
 #   (NB: carry chain input/output must be last
 #        input/output and have been moved there
 #        overriding the alphabetical ordering)
-$__ICE40_FULL_ADDER 1 1 3 2
-589 558 465
-675 609 186 
+$__ICE40_CARRY_WRAPPER 1 1 5 2
+589 558 661 465 465
+675 609 -   -   186
diff --git a/techlibs/ice40/abc_u.box b/techlibs/ice40/abc_u.box
index 42d666051..f00e247b8 100644
--- a/techlibs/ice40/abc_u.box
+++ b/techlibs/ice40/abc_u.box
@@ -3,11 +3,11 @@
 # NB: Inputs/Outputs must be ordered alphabetically
 #     (with exceptions for carry in/out)
 
-# Inputs: A B CI
+# Inputs: A B I0 I3 CI
 # Outputs: O CO
 #   (NB: carry chain input/output must be last
 #        input/output and have been moved there
 #        overriding the alphabetical ordering)
-$__ICE40_FULL_ADDER 1 1 3 2
-1231 1205 874
-675  609  278
+$__ICE40_CARRY_WRAPPER 1 1 5 2
+1231 1205 1285 874 874
+675  609  -    -   278
diff --git a/techlibs/ice40/cells_sim.v b/techlibs/ice40/cells_sim.v
index c7f3bdad2..2a7487f6b 100644
--- a/techlibs/ice40/cells_sim.v
+++ b/techlibs/ice40/cells_sim.v
@@ -142,15 +142,16 @@ module SB_CARRY (output CO, input I0, I1, CI);
 endmodule
 
 (* abc_box_id = 1, lib_whitebox *)
-module \$__ICE40_FULL_ADDER (
+module \$__ICE40_CARRY_WRAPPER (
 	(* abc_carry *)
 	output CO,
 	output O,
-	input A,
-	input B,
+	input A, B,
 	(* abc_carry *)
-	input CI
+	input CI,
+	input I0, I3
 );
+	parameter LUT = 0;
 	SB_CARRY carry (
 		.I0(A),
 		.I1(B),
@@ -158,16 +159,12 @@ module \$__ICE40_FULL_ADDER (
 		.CO(CO)
 	);
 	SB_LUT4 #(
-		//         I0: 1010 1010 1010 1010
-		//         I1: 1100 1100 1100 1100
-		//         I2: 1111 0000 1111 0000
-		//         I3: 1111 1111 0000 0000
-		.LUT_INIT(16'b 0110_1001_1001_0110)
+		.LUT_INIT(LUT)
 	) adder (
-		.I0(1'b0),
+		.I0(I0),
 		.I1(A),
 		.I2(B),
-		.I3(CI),
+		.I3(I3),
 		.O(O)
 	);
 endmodule
diff --git a/techlibs/ice40/ice40_opt.cc b/techlibs/ice40/ice40_opt.cc
index d5106b805..ea56d3f4d 100644
--- a/techlibs/ice40/ice40_opt.cc
+++ b/techlibs/ice40/ice40_opt.cc
@@ -84,7 +84,7 @@ static void run_ice40_opts(Module *module)
 			continue;
 		}
 
-		if (cell->type == "$__ICE40_FULL_ADDER")
+		if (cell->type == "$__ICE40_CARRY_WRAPPER")
 		{
 			SigSpec non_const_inputs, replacement_output;
 			int count_zeros = 0, count_ones = 0;
@@ -114,16 +114,17 @@ static void run_ice40_opts(Module *module)
 				optimized_co.insert(sigmap(cell->getPort("\\CO")[0]));
 				module->connect(cell->getPort("\\CO")[0], replacement_output);
 				module->design->scratchpad_set_bool("opt.did_something", true);
-				log("Optimized $__ICE40_FULL_ADDER cell back to logic (without SB_CARRY) %s.%s: CO=%s\n",
+				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";
-				cell->setPort("\\A", { State::S0, inbit[0], inbit[1], inbit[2] });
+				cell->setPort("\\A", { cell->getPort("\\I0"), inbit[0], inbit[1], cell->getPort("\\I3") });
 				cell->setPort("\\Y", cell->getPort("\\O"));
 				cell->unsetPort("\\B");
 				cell->unsetPort("\\CI");
+				cell->unsetPort("\\I0");
+				cell->unsetPort("\\I3");
 				cell->unsetPort("\\CO");
 				cell->unsetPort("\\O");
-				cell->setParam("\\LUT", RTLIL::Const::from_string("0110100110010110"));
 				cell->setParam("\\WIDTH", 4);
 			}
 			continue;
diff --git a/techlibs/xilinx/cells_sim.v b/techlibs/xilinx/cells_sim.v
index 973e17212..e12b77c02 100644
--- a/techlibs/xilinx/cells_sim.v
+++ b/techlibs/xilinx/cells_sim.v
@@ -380,9 +380,10 @@ endmodule
 
 module SRL16E (
   output Q,
+  input A0, A1, A2, A3, CE,
   (* clkbuf_sink *)
   input CLK,
-  input A0, A1, A2, A3, CE, D
+  input D
 );
   parameter [15:0] INIT = 16'h0000;
   parameter [0:0] IS_CLK_INVERTED = 1'b0;
@@ -401,7 +402,10 @@ endmodule
 module SRLC16E (
   output Q,
   output Q15,
-  input A0, A1, A2, A3, CE, CLK, D
+  input A0, A1, A2, A3, CE,
+  (* clkbuf_sink *)
+  input CLK,
+  input D
 );
   parameter [15:0] INIT = 16'h0000;
   parameter [0:0] IS_CLK_INVERTED = 1'b0;
@@ -422,9 +426,10 @@ module SRLC32E (
   output Q,
   output Q31,
   input [4:0] A,
+  input CE,
   (* clkbuf_sink *)
   input CLK,
-  input CE, D
+  input D
 );
   parameter [31:0] INIT = 32'h00000000;
   parameter [0:0] IS_CLK_INVERTED = 1'b0;
diff --git a/tests/ice40/ice40_opt.ys b/tests/ice40/ice40_opt.ys
new file mode 100644
index 000000000..b17c69c91
--- /dev/null
+++ b/tests/ice40/ice40_opt.ys
@@ -0,0 +1,26 @@
+read_verilog -icells -formal <<EOT
+module top(input CI, I0, output [1:0] CO, output O);
+    wire A = 1'b0, B = 1'b0;
+	\$__ICE40_CARRY_WRAPPER #(
+		//    A[0]: 1010 1010 1010 1010
+		//    A[1]: 1100 1100 1100 1100
+		//    A[2]: 1111 0000 1111 0000
+		//    A[3]: 1111 1111 0000 0000
+		.LUT(~16'b 0110_1001_1001_0110)
+	) u0 (
+		.A(A),
+		.B(B),
+		.CI(CI),
+		.I0(I0),
+		.I3(CI),
+		.CO(CO[0]),
+		.O(O)
+	);
+    SB_CARRY u1 (.I0(~A), .I1(~B), .CI(CI), .CO(CO[1]));
+endmodule
+EOT
+
+equiv_opt -assert -map +/ice40/cells_map.v -map +/ice40/cells_sim.v ice40_opt
+design -load postopt
+select -assert-count 1 t:*
+select -assert-count 1 t:$lut
diff --git a/tests/ice40/run-test.sh b/tests/ice40/run-test.sh
new file mode 100755
index 000000000..ea56b70f0
--- /dev/null
+++ b/tests/ice40/run-test.sh
@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+set -e
+{
+echo "all::"
+for x in *.ys; do
+	echo "all:: run-$x"
+	echo "run-$x:"
+	echo "	@echo 'Running $x..'"
+	echo "	@../../yosys -ql ${x%.ys}.log $x"
+done
+for s in *.sh; do
+	if [ "$s" != "run-test.sh" ]; then
+		echo "all:: run-$s"
+		echo "run-$s:"
+		echo "	@echo 'Running $s..'"
+		echo "	@bash $s"
+	fi
+done
+} > run-test.mk
+exec ${MAKE:-make} -f run-test.mk