Major shifter-ALU redesign

The shifter unit now works in parallel with the ALU and is no longer
part of it. Instructions that use the shifter as input to the ALU will
now take an additional cycle, unless the control unit can detect a
"trivial shift" situation where the shifter's output will be the same as
its input. This change improves Fmax substantially.

1 files changed, 42 insertions, 0 deletions

diff --git a/rtl/core/shifter.sv b/rtl/core/shifter.sv
new file mode 100644
index 0000000..994e76c
--- /dev/null
+++ b/rtl/core/shifter.sv
@@ -0,0 +1,42 @@
+`include "core/uarch.sv"
+
+module core_shifter
+#(parameter W=16)
+(
+	input  shifter_control ctrl,
+	input  logic[W - 1:0]  base,
+	input  logic[7:0]      shift,
+	input  logic           c_in,
+
+	output logic[W - 1:0]  q,
+	output logic           c
+);
+
+	localparam LOG = $clog2(W);
+
+	logic[W - 1:0] q_no_c, q_shl, q_shr, q_ror;
+	logic[W:0] sign_mask;
+	logic c_shl, c_shr, c_ror;
+
+	assign sign_mask = {(W + 1){ctrl.sign_extend & base[W - 1]}};
+	assign {c_shl, q_shl} = {c_in, base} << shift;
+	assign {q_shr, c_shr} = {base, c_in} >> shift | ~(sign_mask >> shift);
+
+	logic ror_cycle;
+	logic[LOG - 1:0] ror_shift;
+	logic[2 * W:0] ror_out;
+
+	assign ror_shift = shift[LOG - 1:0];
+	assign ror_cycle = |shift[7:LOG] & ~|ror_shift;
+	assign ror_out = {base, base, c_in} >> {ror_cycle, ror_shift};
+	assign {q_ror, c_ror} = ror_out[W:0];
+
+	always_comb
+		if(ctrl.ror)
+			{c, q} = {c_ror, q_ror};
+		else if(ctrl.shr)
+			{c, q} = {c_shr, q_shr[W - 1] | (ctrl.put_carry & c_in), q_shr[W - 2:0]};
+		else
+			{c, q} = {c_shl, q_shl};
+
+endmodule