diff --git a/passes/pmgen/xilinx_dsp.cc b/passes/pmgen/xilinx_dsp.cc
index 1732a2d6a..b3d302071 100644
--- a/passes/pmgen/xilinx_dsp.cc
+++ b/passes/pmgen/xilinx_dsp.cc
@@ -40,6 +40,7 @@ void pack_xilinx_dsp(dict<SigBit, Cell*> &bit_to_driver, xilinx_dsp_pm &pm)
 	log("dsp:     %s\n", log_id(st.dsp, "--"));
 	log("ffM:     %s\n", log_id(st.ffM, "--"));
 	log("addAB:   %s\n", log_id(st.addAB, "--"));
+	log("muxAB:   %s\n", log_id(st.muxAB, "--"));
 	log("ffP:     %s\n", log_id(st.ffP, "--"));
 	//log("muxP:  %s\n", log_id(st.muxP, "--"));
 	log("sigPused: %s\n", log_signal(st.sigPused));
@@ -58,7 +59,11 @@ void pack_xilinx_dsp(dict<SigBit, Cell*> &bit_to_driver, xilinx_dsp_pm &pm)
 		log("  adder %s (%s)\n", log_id(st.addAB), log_id(st.addAB->type));
 
 		SigSpec &opmode = cell->connections_.at("\\OPMODE");
-		if (st.ffP && C == P) {
+		if (st.ffP && st.muxAB) {
+			opmode[4] = st.muxAB->getPort("\\S");
+			pm.autoremove(st.muxAB);
+		}
+		else if (st.ffP && C == P) {
 			C = SigSpec();
 			opmode[4] = State::S0;
 		}
diff --git a/passes/pmgen/xilinx_dsp.pmg b/passes/pmgen/xilinx_dsp.pmg
index 132b09b2b..fdc3fa5e7 100644
--- a/passes/pmgen/xilinx_dsp.pmg
+++ b/passes/pmgen/xilinx_dsp.pmg
@@ -3,7 +3,7 @@ pattern xilinx_dsp
 state <SigBit> clock
 state <std::set<SigBit>> sigAset sigBset
 state <SigSpec> sigC sigM sigMused sigP sigPused
-state <Cell*> addAB
+state <Cell*> addAB muxAB
 
 match dsp
 	select dsp->type.in(\DSP48E1)
@@ -172,34 +172,7 @@ match ffP
 	optional
 endmatch
 
-//// $mux cell left behind by dff2dffe
-////   would prefer not to run 'opt_expr -mux_undef'
-////   since that would lose information helpful for
-////   efficient wide-mux inference
-//match muxP
-//	if !sigPused.empty() && !ffP
-//	select muxP->type.in($mux)
-//	select nusers(port(muxP, \B)) == 2
-//	select port(muxP, \A).is_fully_undef()
-//	filter param(muxP, \WIDTH).as_int() >= GetSize(sigPused)
-//	filter includes(port(muxP, \B).to_sigbit_set(), sigPused.to_sigbit_set())
-//	optional
-//endmatch
-//
-//match ffY
-//	if muxP
-//	select ffY->type.in($dff, $dffe)
-//	select nusers(port(ffY, \D)) == 2
-//	// DSP48E1 does not support clock inversion
-//	select param(ffY, \CLK_POLARITY).as_bool()
-//	filter param(ffY, \WIDTH).as_int() >= GetSize(sigPused)
-//	filter includes(port(ffY, \D).to_sigbit_set(), port(muxP, \Y).to_sigbit_set())
-//endmatch
-
-code ffP clock
-//	if (ffY)
-//		ffP = ffY;
-
+code ffP sigP clock
 	if (ffP) {
 		for (auto b : port(ffP, \Q))
 			if (b.wire->get_bool_attribute(\keep))
@@ -211,7 +184,46 @@ code ffP clock
 			reject;
 
 		clock = c;
-	}
 
+		sigP = port(ffP, \Q);
+	}
+endcode
+
+match muxA
+	if addAB
+	select muxA->type.in($mux)
+	select nusers(port(muxA, \Y)) == 2
+	index <SigSpec> port(muxA, \A) === sigP
+	index <SigSpec> port(muxA, \Y) === sigC
+	optional
+endmatch
+
+match muxB
+	if addAB
+	select muxB->type.in($mux)
+	select nusers(port(muxB, \Y)) == 2
+	index <SigSpec> port(muxB, \B) === sigP
+	index <SigSpec> port(muxB, \Y) === sigC
+	optional
+endmatch
+
+code sigC muxAB
+	if (muxA) {
+		muxAB = muxA;
+		sigC = port(muxAB, \B);
+	}
+	if (muxB) {
+		muxAB = muxB;
+		sigC = port(muxAB, \A);
+	}
+	if (muxAB) {
+		// Ensure that adder is not used
+		SigSpec opmodeZ = port(dsp, \OPMODE).extract(4,3);
+		if (!opmodeZ.is_fully_zero())
+			reject;
+	}
+endcode
+
+code
 	accept;
 endcode