Fetch: Allow new queued jumps to take precedence over old queued jumps (i.e., ldm...

[firearm.git] / Execute.v

diff --git a/Execute.v b/Execute.v
index 7f4cf6c988df5ac3c647bd12c6509c67b004f62c..7dcc4b991d7da72df4adf9b34bfcc1a7ee61af43 100644 (file)
--- a/Execute.v
+++ b/Execute.v
@@ -19,6 +19,7 @@ module Execute(
        output reg outbubble = 1,
        output reg [31:0] outcpsr = 0,
        output reg [31:0] outspsr = 0,
+       output reg outcpsrup = 0,
        output reg write_reg = 1'bx,
        output reg [3:0] write_num = 4'bxxxx,
        output reg [31:0] write_data = 32'hxxxxxxxx,
@@ -42,6 +43,7 @@ module Execute(
        
        reg next_outbubble;
        reg [31:0] next_outcpsr, next_outspsr;
+       reg next_outcpsrup;
        reg next_write_reg;
        reg [3:0] next_write_num;
 
@@ -57,7 +59,7 @@ module Execute(
                .in0(alu_in0), .in1(alu_in1), .cpsr(cpsr), .op(alu_op),
                .setflags(alu_setflags), .shifter_carry(carry),
                .result(alu_result), .cpsr_out(alu_outcpsr), .setres(alu_setres));
-       
+
        always @(posedge clk)
        begin
                if (!stall)
@@ -65,6 +67,7 @@ module Execute(
                        outbubble <= next_outbubble;
                        outcpsr <= next_outcpsr;
                        outspsr <= next_outspsr;
+                       outcpsrup <= next_outcpsrup;
                        write_reg <= next_write_reg;
                        write_num <= next_write_num;
                        write_data <= next_write_data;
@@ -75,6 +78,13 @@ module Execute(
                        outop2 <= op2;
                end
        end
+       
+       reg delayedflush = 0;
+       always @(posedge clk)
+               if (flush && outstall /* halp! I can't do it now, maybe later? */)
+                       delayedflush <= 1;
+               else if (!outstall /* anything has been handled this time around */)
+                       delayedflush <= 0;
 
        reg prevstall = 0;
        always @(posedge clk)
@@ -83,9 +93,10 @@ module Execute(
        always @(*)
        begin
                outstall = stall;
-               next_outbubble = inbubble | flush;
+               next_outbubble = inbubble | flush | delayedflush;
                next_outcpsr = cpsr;
                next_outspsr = spsr;
+               next_outcpsrup = 0;
                next_write_reg = 0;
                next_write_num = 4'hx;
                next_write_data = 32'hxxxxxxxx;
@@ -101,7 +112,7 @@ module Execute(
                alu_setflags = 1'bx;
 
                jmp = 1'b0;
-               jmppc = 32'hxxxxxxxx;
+               jmppc = 32'h00000000;
 
                casez (insn)
                `DECODE_ALU_MULT:       /* Multiply -- must come before ALU, because it pattern matches a specific case of ALU */
@@ -114,9 +125,10 @@ module Execute(
                                mult_in1 = op2 /* Rs */;
                                $display("New MUL instruction");
                        end
-                       outstall = stall | ((!prevstall | !mult_done) && !inbubble);
-                       next_outbubble = inbubble | !mult_done | !prevstall;
+                       outstall = outstall | ((!prevstall | !mult_done) && !inbubble);
+                       next_outbubble = next_outbubble | !mult_done | !prevstall;
                        next_outcpsr = insn[20] /* S */ ? {mult_result[31] /* N */, mult_result == 0 /* Z */, 1'b0 /* C */, cpsr[28] /* V */, cpsr[27:0]} : cpsr;
+                       next_outcpsrup = insn[20] /* S */;
                        next_write_reg = 1;
                        next_write_num = insn[19:16] /* Rd -- why the fuck isn't this the same place as ALU */;
                        next_write_data = mult_result;
@@ -130,9 +142,11 @@ module Execute(
                                next_write_data = spsr;
                        else
                                next_write_data = cpsr;
+                       next_outcpsrup = 1;
                end
                `DECODE_ALU_MSR,        /* MSR (Transfer register to PSR) */
                `DECODE_ALU_MSR_FLAGS:  /* MSR (Transfer register or immediate to PSR, flag bits only) */
+               begin
                        if ((cpsr[4:0] == `MODE_USR) || (insn[16] /* that random bit */ == 1'b0))       /* flags only */
                        begin
                                if (insn[22] /* Ps */)
@@ -145,6 +159,8 @@ module Execute(
                                else
                                        next_outcpsr = op0;
                        end
+                       next_outcpsrup = 1;
+               end
                `DECODE_ALU_SWP,        /* Atomic swap */
                `DECODE_ALU_BX,         /* Branch */
                `DECODE_ALU_HDATA_REG,  /* Halfword transfer - register offset */
@@ -164,6 +180,7 @@ module Execute(
                        end
                        
                        next_outcpsr = ((insn[15:12] == 4'b1111) && insn[20]) ? spsr : alu_outcpsr;
+                       next_outcpsrup = insn[20] /* S */;
                end
                `DECODE_LDRSTR_UNDEFINED,       /* Undefined. I hate ARM */
                `DECODE_LDRSTR,         /* Single data transfer */
@@ -171,12 +188,12 @@ module Execute(
                begin end
                `DECODE_BRANCH:
                begin
-                       if(!inbubble) begin
+                       if(!inbubble && !flush && !delayedflush) begin
                                jmppc = pc + op0 + 32'h8;
                                if(insn[24]) begin
                                        next_write_reg = 1;
                                        next_write_num = 4'hE; /* link register */
-                                       next_write_data = pc - 32'h4;
+                                       next_write_data = pc + 32'h4;
                                end
                                jmp = 1'b1;
                        end
@@ -251,7 +268,7 @@ module ALU(
 
        assign sum = {1'b0, in0} + {1'b0, in1};
        assign diff = {1'b0, in0} - {1'b0, in1};
-       assign rdiff = {1'b0, in1} + {1'b0, in0};
+       assign rdiff = {1'b0, in1} - {1'b0, in0};
        assign sum_v = (in0[31] ^~ in1[31]) & (sum[31] ^ in0[31]);
        assign diff_v = (in0[31] ^ in1[31]) & (diff[31] ^ in0[31]);
        assign rdiff_v = (in0[31] ^ in1[31]) & (rdiff[31] ^ in1[31]);