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.

2 files changed, 11 insertions, 63 deletions

diff --git a/rtl/core/alu/alu.sv b/rtl/core/alu/alu.sv
index 6ded727..d999164 100644
--- a/rtl/core/alu/alu.sv
+++ b/rtl/core/alu/alu.sv
@@ -3,10 +3,9 @@
 module core_alu
 #(parameter W=16)
 (
-	input  alu_control    ctrl,
+	input  alu_op         op,
 	input  logic[W - 1:0] a,
-	                      base,
-	input  logic[7:0]     shift,
+	                      b,
 	input  logic          c_in,
 
 	output logic[W - 1:0] q,
@@ -14,22 +13,13 @@ module core_alu
 	output logic          v_valid
 );
 
-	logic c, v, swap, sub, and_not, c_shifter, c_add, v_add;
-	logic[W - 1:0] b, swap_a, swap_b, not_b, c_in_add, q_add, q_and, q_orr, q_xor;
+	logic c, v, swap, sub, and_not, c_add, v_add;
+	logic[W - 1:0] swap_a, swap_b, not_b, c_in_add, q_add, q_and, q_orr, q_xor;
 
 	assign swap_a = swap ? b : a;
 	assign swap_b = swap ? a : b;
 	assign not_b = ~b;
 
-	core_alu_shifter #(.W(W)) shifter
-	(
-		.base(base),
-		.shift(shift),
-		.b(b),
-		.c(c_shifter),
-		.*
-	);
-
 	core_alu_add #(.W(W)) op_add
 	(
 		.a(swap_a),
@@ -61,7 +51,7 @@ module core_alu
 	);
 
 	always_comb begin
-		unique case(ctrl.op)
+		unique case(op)
 			`ALU_ADD, `ALU_ADC, `ALU_CMN, `ALU_CMP, `ALU_SUB, `ALU_SBC:
 				swap = 0;
 
@@ -72,7 +62,7 @@ module core_alu
 				swap = 1'bx;
 		endcase
 
-		unique case(ctrl.op)
+		unique case(op)
 			`ALU_ADD, `ALU_CMN, `ALU_ADC:
 				sub = 0;
 
@@ -83,7 +73,7 @@ module core_alu
 				sub = 1'bx;
 		endcase
 
-		unique case(ctrl.op)
+		unique case(op)
 			`ALU_ADD, `ALU_CMN, `ALU_CMP, `ALU_SUB, `ALU_RSB:
 				c_in_add = 0;
 
@@ -97,7 +87,7 @@ module core_alu
 				c_in_add = {W{1'bx}};
 		endcase
 
-		unique case(ctrl.op)
+		unique case(op)
 			`ALU_AND, `ALU_TST:
 				and_not = 0;
 
@@ -108,7 +98,7 @@ module core_alu
 				and_not = 1'bx;
 		endcase
 
-		unique case(ctrl.op)
+		unique case(op)
 			`ALU_SUB, `ALU_RSB, `ALU_ADD, `ALU_ADC, `ALU_SBC, `ALU_RSC, `ALU_CMP, `ALU_CMN:
 				q = q_add;
 
@@ -128,9 +118,9 @@ module core_alu
 				q = not_b;
 		endcase
 
-		unique case(ctrl.op)
+		unique case(op)
 			`ALU_AND, `ALU_EOR, `ALU_TST, `ALU_TEQ, `ALU_ORR, `ALU_MOV, `ALU_BIC, `ALU_MVN: begin
-				c = c_shifter;
+				c = c_in;
 				v = 1'bx;
 				v_valid = 0;
 			end
diff --git a/rtl/core/alu/shifter.sv b/rtl/core/alu/shifter.sv
deleted file mode 100644
index 1f99e99..0000000
--- a/rtl/core/alu/shifter.sv
+++ /dev/null
@@ -1,42 +0,0 @@
-`include "core/uarch.sv"
-
-module core_alu_shifter
-#(parameter W=16)
-(
-	input  alu_control    ctrl,
-	input  logic[W - 1:0] base,
-	input  logic[7:0]     shift,
-	input  logic          c_in,
-
-	output logic[W - 1:0] b,
-	output logic          c
-);
-
-	localparam LOG = $clog2(W);
-
-	logic[W - 1:0] b_no_c, b_shl, b_shr, b_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, b_shl} = {c_in, base} << shift;
-	assign {b_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 {b_ror, c_ror} = ror_out[W:0];
-
-	always_comb
-		if(ctrl.ror)
-			{c, b} = {c_ror, b_ror};
-		else if(ctrl.shr)
-			{c, b} = {c_shr, b_shr[W - 1] | (ctrl.put_carry & c_in), b_shr[W - 2:0]};
-		else
-			{c, b} = {c_shl, b_shl};
-
-endmodule