diff --git a/passes/proc/proc_mux.cc b/passes/proc/proc_mux.cc
index 68127d1bc..b60b4fff2 100644
--- a/passes/proc/proc_mux.cc
+++ b/passes/proc/proc_mux.cc
@@ -393,6 +393,27 @@ RTLIL::SigSpec signal_to_mux_tree(RTLIL::Module *mod, SnippetSwCache &swcache, d
 			if (full_case_bits.count(sig[i]))
 				result[i] = State::Sx;
 
+		// For full_case switches, if a majority of arms assign the same value to a bit
+		// (via direct actions), use that as the else-branch seed instead of Sx. This
+		// lets proc_mux skip generating mux cells for arms that produce the dominant value.
+		if (!full_case_bits.empty() && !sw->cases.empty()) {
+			int n = GetSize(sw->cases);
+			std::vector<dict<SigBit, int>> counts(GetSize(sig));
+			for (auto cs2 : sw->cases) {
+				RTLIL::SigSpec probe = RTLIL::SigSpec(RTLIL::State::Sx, sig.size());
+				for (auto &action : cs2->actions)
+					sig.replace(action.first, action.second, &probe);
+				for (int i = 0; i < GetSize(sig); i++)
+					if (full_case_bits.count(sig[i]) && probe[i] != RTLIL::State::Sx)
+						counts[i][probe[i]]++;
+			}
+			for (int i = 0; i < GetSize(sig); i++) {
+				if (!full_case_bits.count(sig[i])) continue;
+				for (auto &kv : counts[i])
+					if (kv.second * 2 > n) { result[i] = kv.first; break; }
+			}
+		}
+
 		// evaluate in reverse order to give the first entry the top priority
 		RTLIL::SigSpec initial_val = result;
 		RTLIL::Cell *last_mux_cell = NULL;
diff --git a/tests/proc/proc_mux_dominant.v b/tests/proc/proc_mux_dominant.v
new file mode 100644
index 000000000..1e02533b2
--- /dev/null
+++ b/tests/proc/proc_mux_dominant.v
@@ -0,0 +1,50 @@
+// Test cases for proc_mux dominant-value optimization.
+//
+// When a full_case switch has a majority of arms assigning the same value to a
+// signal bit, proc_mux uses that dominant value as the starting point instead
+// of Sx.  Arms that produce the dominant value are then skipped (the mux
+// condition evaluates to "when == else"), avoiding spurious $eq/$mux cells.
+
+// dominant_explicit: 3 of 4 arms assign the same constants (dominant values).
+// Expected after proc: one $mux per output word, one $logic_not for the
+// selector — zero $eq cells, zero $pmux cells.
+module dominant_explicit(input [1:0] s, output reg [2:0] y, output reg [1:0] z);
+  always @* begin
+    y = 3'b001;
+    z = 2'b00;
+    case (s)
+      2'b00: begin y = 3'b110; z = 2'b11; end  // only arm that differs
+      2'b01: begin y = 3'b001; z = 2'b00; end  // explicit dominant
+      2'b10: begin y = 3'b001; z = 2'b00; end  // explicit dominant
+      2'b11: begin y = 3'b001; z = 2'b00; end  // explicit dominant
+    endcase
+  end
+endmodule
+
+// dominant_wire: dominant value is an input wire (not a constant).
+// Expected after proc: 1 $logic_not + 1 $mux, no $eq/$pmux.
+module dominant_wire(input [1:0] s, input [2:0] a, output reg [2:0] y);
+  always @* begin
+    y = a;
+    case (s)
+      2'b00: y = 3'b110;  // only arm that differs
+      2'b01: y = a;       // explicit dominant
+      2'b10: y = a;       // explicit dominant
+      2'b11: y = a;       // explicit dominant
+    endcase
+  end
+endmodule
+
+// no_dominant: all four arms assign distinct values — no majority.
+// The optimization must NOT fire; behavior must be unchanged.
+// Expected after proc: $eq cells for each non-zero compare arm, $pmux.
+module no_dominant(input [1:0] s, input [2:0] a, b, c, d, output reg [2:0] y);
+  always @* begin
+    case (s)
+      2'b00: y = a;
+      2'b01: y = b;
+      2'b10: y = c;
+      2'b11: y = d;
+    endcase
+  end
+endmodule
diff --git a/tests/proc/proc_mux_dominant.ys b/tests/proc/proc_mux_dominant.ys
new file mode 100644
index 000000000..7cdce9178
--- /dev/null
+++ b/tests/proc/proc_mux_dominant.ys
@@ -0,0 +1,40 @@
+# Test that proc_mux uses a dominant-value pre-scan to avoid generating
+# unnecessary mux cells when a full_case switch has a majority of arms
+# assigning the same value.
+
+# 3 of 4 arms assign identical constants for both outputs.
+# The optimization should seed the result with the dominant values (3'b001,
+# 2'b00) so only the one differing arm (2'b00 -> 3'b110, 2'b11) generates
+# cells.  Each output word is a separate SigSnippet, so we expect one
+# $logic_not + one $mux per word = 2 of each total.
+read_verilog proc_mux_dominant.v
+hierarchy -top dominant_explicit
+proc
+select -assert-count 2 t:$logic_not
+select -assert-count 2 t:$mux
+select -assert-count 0 t:$eq
+select -assert-count 0 t:$pmux
+
+# 3 of 4 arms pass through an input wire 'a'.  The dominant value is a wire
+# signal rather than a constant; the optimization must still recognise it.
+# Only one arm differs (2'b00 -> 3'b110), producing 1 $logic_not + 1 $mux.
+design -reset
+read_verilog proc_mux_dominant.v
+hierarchy -top dominant_wire
+proc
+select -assert-count 1 t:$logic_not
+select -assert-count 1 t:$mux
+select -assert-count 0 t:$eq
+select -assert-count 0 t:$pmux
+
+# All four arms assign distinct values; no majority exists.  The optimization
+# must not fire and the generated netlist must be functionally correct.
+design -reset
+read_verilog proc_mux_dominant.v
+hierarchy -top no_dominant
+proc
+# Three explicit non-zero compare arms each produce an $eq; the 2'b00 arm
+# uses $logic_not (all-zero check); all results are merged into one $pmux.
+select -assert-count 3 t:$eq
+select -assert-count 1 t:$logic_not
+select -assert-count 1 t:$pmux