%% LyX 1.3 created this file. For more info, see http://www.lyx.org/.
%% Do not edit unless you really know what you are doing.
\documentclass[twocolumn,american]{article}
\usepackage{ae}
\usepackage{aecompl}
\usepackage[T1]{fontenc}
\usepackage[latin1]{inputenc}
\usepackage{a4wide}
\usepackage{geometry}
\geometry{verbose,a4paper,lmargin=0.5in,rmargin=0.5in}
\usepackage{fancyhdr}
\pagestyle{fancy}
\usepackage{graphicx}
\IfFileExists{url.sty}{\usepackage{url}}
{\newcommand{\url}{\texttt}}
\makeatletter
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% LyX specific LaTeX commands.
\newcommand{\noun}[1]{\textsc{#1}}
%% Bold symbol macro for standard LaTeX users
\newcommand{\boldsymbol}[1]{\mbox{\boldmath $#1$}}
%% Because html converters don't know tabularnewline
\providecommand{\tabularnewline}{\\}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% Textclass specific LaTeX commands.
\usepackage{noweb}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% User specified LaTeX commands.
\usepackage[dvips,colorlinks=true,linkcolor=blue]{hyperref}
\usepackage{babel}
\makeatother
\begin{document}
\title{b16 --- A Forth Processor in an FPGA}
\author{\noun{Bernd Paysan}}
\maketitle
\lhead{b16 --- A Forth Processor in an FPGA}\chead{\noun{Bernd
Paysan}}
\begin{abstract}
This article presents architecture and implementation of the b16 stack
processor. This processor is inspired by \noun{Chuck Moore'}s newest
Forth processors. The minimalistic design fits into small FPGAs and
ASICs and is ideally suited for applications that need both control
and calculations. The synthesizible implementation uses Verilog.
\end{abstract}
\section*{Introduction}
Minimalistic CPUs can be used in many designs. A state machine often
is too complicated and too difficult to develop, when there are more
than a few states. A program with subroutines can perform a lot more
complex tasks, and is easier to develop at the same time. Also, ROM-
and RAM blocks occupy much less place on silicon than {}``random
logic''. That's also valid for FPGAs, where {}``block RAM'' is
--- in contrast to logic elements --- plenty.
The architecture is inspired by the c18 from \noun{Chuck Moore}
\cite{c18}. The exact instruction mix is different. I traded \texttt{2{*}}
and \texttt{2/} against division step and Forth-typical logic operations;
these two instructions can be implemented as short macro. Also, this
architecture is byte-addressed.
The original concept (which was synthesizible, and could execute a
small sample program) was written in an afternoon. The current version
is somewhat faster, and really runs on a Altera Flex10K30E on a FPGA
evaluation board from \noun{Hans Eckes}. Size and speed of the processor
can be evaluated.
\begin{description}
\item [Flex10K30E]About 600 LCs, the unit for logic cells in Altera%
\footnote{A logic cell can compute a logic function with four inputs and one
output, or a full-adder, and also contains a flip-flop.%
}. The logic to interface with the eval board needs another 100 LCs.
The slowest model runs at up to 25MHz.
\item [Xfab~0.6\ensuremath{稀]$\sim$1mm\ensuremath{瓠 with 8 stack elements,
that's a technology with only 2 metal layers.
\item [TSMC~0.5\ensuremath{稀]$<$0.4mm\ensuremath{瓠 with 8 stack elements,
this technology has 3 metal layers. With a somewhat optimized ALU
the 5V library reaches 100MHz.
\end{description}
The complete development (excluding board layout and test synthesis
for ASIC processes) was done with free or zero cost tools. Icarus
Verilog in the current version is quite useful for projects in this
order of magnitude, and Quartus II Web Edition is a big chunk to download,
but doesn't cost anything (downside: Windows NT, the version for real
operating system costs real money).
A word about Verilog: Verilog is a C-like language, but tailored for
the purpose to simulate logic, and to write synthesizible code. Variables
are bits and bit vectors, and assignments are typically non-blocking,
i.e. on assignments first all right sides are computed, and the left
sides are modified afterwards. Also, Verilog has events, like changing
of values or clock edges, and blocks can wait on them.
\section{Architectural Overview}
The core components are
\begin{itemize}
\item An ALU
\item A data stack with top and next of stack (T and N) as inputs for the
ALU
\item A return stack, where the top of return stack (R) can be used as address
\item An instruction pointer P
\item An address register A
\item An address latch \texttt{addr}, to address external memory
\item An instruction latch I
\end{itemize}
Figure \ref{blockdiagram} shows a block diagram.
%
\begin{figure}
\begin{center}\includegraphics[%
width=1.0\columnwidth]{b16.eps}\end{center}
\caption{Block Diagram\label{blockdiagram}}
\end{figure}
\subsection{Register}
In addition to the user-visible latches there are control latches
for external RAM (\texttt{rd} and \texttt{wr}), stack pointers (\texttt{sp}
and \texttt{rp}), a carry \texttt{c} and the flag \texttt{incby},
by which \texttt{addr} is incremented.
\medskip{}
\begin{center}\begin{tabular}{|c|l|}
\hline
\emph{Name}&
\emph{Function}\tabularnewline
\hline
\hline
T&
Top of Stack\tabularnewline
\hline
N&
Next of Stack\tabularnewline
\hline
I&
Instruction Bundle\tabularnewline
\hline
P&
Program Counter\tabularnewline
\hline
A&
Address Register\tabularnewline
\hline
addr&
Address Latch\tabularnewline
\hline
state&
Processor State\tabularnewline
\hline
sp&
Stack Pointer\tabularnewline
\hline
rp&
Return Stack Pointer\tabularnewline
\hline
c&
Carry Flag\tabularnewline
\hline
incby&
Increment Address by byte/word\tabularnewline
\hline
\end{tabular}\end{center}
\medskip{}
<<register declarations>>=
reg rd;
reg [1:0] wr;
reg [sdep-1:0] sp;
reg [rdep-1:0] rp;
reg `L T, N, I, P, A, addr;
reg [2:0] state;
reg c;
reg incby;
reg intack;
@
\section{Instruction Set}
There are 32 different instructions. Since several instructions fit
into a 16 bit word, we call the bits to store the packed instructions
in an instruction word {}``slot'', and the instruction word itself
{}``bundle''. The arrangement here is 1,5,5,5, i.e. the first slot
is only one bit large (the more significant bits are filled with 0),
and the others all 5 bits.\filbreak
The operations in one instruction word are executed one after the
other. Each instruction takes one cycle, memory operation (including
instruction fetch) need another cycle. Which instruction is to be
executed is stored in the variable \texttt{state}.\filbreak
The instruction set is divided into four groups: jumps, ALU, memory,
and stack. Table \ref{instructions} shows an overview over the instruction
set.%
\begin{table*}
\begin{center}\begin{tabular}{|c|c|c|c|c|c|c|c|c|l|}
\hline
&
0 &
1 &
2 &
3 &
4 &
5 &
6 &
7 &
\emph{Comment}\tabularnewline
\hline
0 &
nop&
call&
jmp&
ret&
jz&
jnz&
jc&
jnc&
\tabularnewline
&
&
exec&
goto&
ret&
gz&
gnz&
gc&
gnc&
\emph{for slot 3 }\tabularnewline
\hline
8 &
xor&
com&
and&
or&
+&
+c&
$*+$&
/--&
\tabularnewline
\hline
10 &
A!+&
A@+&
R@+&
lit&
Ac!+&
Ac@+&
Rc@+&
litc&
\tabularnewline
&
A!&
A@&
R@&
lit&
Ac!&
Ac@&
Rc@&
litc&
\emph{for slot 1} \tabularnewline
\hline
18 &
nip&
drop&
over&
dup&
>r&
>a&
r>&
a&
\tabularnewline
\hline
\end{tabular}\end{center}
\caption{Instruction Set\label{instructions}}
\end{table*}
\filbreak
Jumps use the rest of the instruction word as target address (except
\texttt{ret}). The lower bits of the instruction pointer P are replaced,
there's nothing added. For instructions in the last slot, no address
remains, so they use T (TOS) as target.\filbreak
<<instruction selection>>=
// instruction and branch target selection
reg [4:0] inst;
reg `L jmp;
always @(state or I)
case(state[1:0])
2'b00: inst <= { 4'b0000, I[15] };
2'b01: inst <= I[14:10];
2'b10: inst <= I[9:5];
2'b11: inst <= I[4:0];
endcase // casez(state)
always @(state or I or P or T)
case(state[1:0])
2'b00: jmp <= { I[14:0], 1'b0 };
2'b01: jmp <= { P[15:11], I[9:0], 1'b0 };
2'b10: jmp <= { P[15:6], I[4:0], 1'b0 };
2'b11: jmp <= { T[15:1], 1'b0 };
endcase // casez(state)
@
The instructions themselves are executed depending on \texttt{inst}:
<<instructions>>=
casez(inst)
<<control flow>>
<<ALU operations>>
<<load/store>>
<<stack operations>>
endcase // case(inst)
@
\subsection{Jumps}
In detail, jumps are performed as follows: the target address is stored
in the address latch \texttt{addr}, which addresses memory, not in
the P register. The register P will be set to the incremented value
of \texttt{addr}, after the instruction fetch cycle. Apart from \texttt{call},
\texttt{jmp} and \texttt{ret} there are conditional jumps, which test
for 0 and carry. The lowest bit of the return stack is used to save
the carry flag across calls. Conditional instructions don't consume
the tested value, which is different from Forth.\filbreak
To make it easier to understand, I also define the effect of an instruction
in a pseudo language:\filbreak
\begin{description}
\item [nop]( --- )\filbreak
\item [call]( --- r:P ) $\mathrm{P}\leftarrow jmp$; $\mathrm{c}\leftarrow0$
\filbreak
\item [jmp]( --- ) $\mathrm{P}\leftarrow jmp$\filbreak
\item [ret]( r:a --- ) $\mathrm{P}\leftarrow a\wedge\$\mathrm{FFFE}$;
$\mathrm{c}\leftarrow a\wedge1$ \filbreak
\item [jz]( n --- n ) $\mathbf{if}(n=0)\,\mathrm{P}\leftarrow jmp$ \filbreak
\item [jnz]( n --- n ) $\mathbf{if}(n\ne0)\,\mathrm{P}\leftarrow jmp$
\filbreak
\item [jc]( --- ) $\mathbf{if}(c)\,\mathrm{P}\leftarrow jmp$ \filbreak
\item [jnc]( --- ) $\mathbf{if}(c=0)\,\mathrm{P}\leftarrow jmp$ \filbreak
\end{description}
<<control flow>>=
5'b00001: begin
rp <= rpdec;
addr <= jmp;
c <= 1'b0;
if(state == 3'b011) `DROP;
end // case: 5'b00001
5'b00010: begin
addr <= jmp;
if(state == 3'b011) `DROP;
end
5'b00011: begin
{ c, addr } <= { R[0], R[l-1:1], 1'b0 };
rp <= rpinc;
end // case: 5'b01111
5'b001??: begin
if((inst[1] ? c : zero) ^ inst[0])
addr <= jmp;
if(state == 3'b011) `DROP;
end
@
\subsection{ALU Operations}
The ALU instructions use the ALU, which computes a result \texttt{res}
and a carry bit from T and N. The instruction \texttt{com} is an exception,
since it only inverts T --- that doesn't require an ALU.\filbreak
The two instructions \texttt{{*}+} (multiplication step) and \texttt{/-}
(division step) shift the result into the A register and carry bit. \texttt{{*}+}
adds N to T, when the carry bit is set, and shifts the result one
step right.\filbreak
\texttt{/-} also adds N to T, but also tests, if there is an overflow,
or if the old carry was set. The result is shifted one to the left.\filbreak
Ordinary ALU instructions just write the result of the ALU into T
and c, and reload N.\filbreak
\begin{description}
\item [xor]( a b --- r ) $r\leftarrow a\oplus b$\filbreak
\item [com]( a --- r ) $r\leftarrow a\oplus\$\mathrm{FFFF}$, $\mathrm{c}\leftarrow1$\filbreak
\item [and]( a b --- r ) $r\leftarrow a\wedge b$\filbreak
\item [or]( a b --- r ) $r\leftarrow a\vee b$\filbreak
\item [+]( a b --- r ) $\mathrm{c},r\leftarrow a+b$\filbreak
\item [+c]( a b --- r) $\mathrm{c},r\leftarrow a+b+\mathrm{c}$\filbreak
\item [$*$+]( a b --- a r ) $\mathbf{if}(\mathrm{c})\, c_{n},r\leftarrow a+b\,\mathbf{else}\, c_{n},r\leftarrow0,b$;
$r,\mathrm{A},\mathrm{c}\leftarrow c_{n},r,\mathrm{A}$\filbreak
\item [/--]( a b --- a r ) $c_{n},r_{n}\leftarrow a+b+1;$ $\mathbf{if}(\mathrm{c}\vee c_{n})\, r\leftarrow r_{n}$;
$\mathrm{c},r,\mathrm{A}\leftarrow r,\mathrm{A},\mathrm{c}\vee c_{n}$
\filbreak
\end{description}
<<ALU operations>>=
5'b01001: { c, T } <= { 1'b1, ~T };
5'b01110: { T, A, c } <=
{ c ? { carry, res } : { 1'b0, T }, A };
5'b01111: { c, T, A } <=
{ (c | carry) ? res : T, A, (c | carry) };
5'b01???: begin
c <= carry;
{ sp, T, N } <= { spinc, res, toN };
end // case: 5'b01???
@
\subsection{Memory Instructions}
\noun{Chuck Moore} doesn't use the TOS as address any more, but
has introduced the A register. When you want to copy memory areas,
you need a second address register, that's what he uses the top of
return stack R for. Since P has to be incremented after each instruction
fetch (to point to the next instruction), the address logic must have
auto increment. This will also be used for other accesses.\filbreak
Memory instructions which use the first slot, and don't index over
P, don't increment the pointer; that's to realize read-modify-write
instructions like \texttt{+!}. Write access is only possible via A,
the two other pointers can only be used for read access.\filbreak
\begin{description}
\item [A!+]( n --- ) $mem[\mathrm{A}]\leftarrow n$; $\mathrm{A}\leftarrow\mathrm{A}+2$\filbreak
\item [A@+]( --- n ) $n\leftarrow mem[\mathrm{A}]$; $\mathrm{A}\leftarrow\mathrm{A}+2$\filbreak
\item [R@+]( --- n ) $n\leftarrow mem[\mathrm{R}]$; $\mathrm{R}\leftarrow\mathrm{R}+2$\filbreak
\item [lit]( --- n ) $n\leftarrow mem[\mathrm{P}]$; $\mathrm{P}\leftarrow\mathrm{P}+2$\filbreak
\item [Ac!+]( c --- ) $mem.b[\mathrm{A}]\leftarrow c$; $\mathrm{A}\leftarrow\mathrm{A}+1$\filbreak
\item [Ac@+]( --- c ) $c\leftarrow mem.b[\mathrm{A}]$; $\mathrm{A}\leftarrow\mathrm{A}+1$\filbreak
\item [Rc@+]( --- c ) $c\leftarrow mem.b[\mathrm{R}]$; $\mathrm{R}\leftarrow\mathrm{R}+1$\filbreak
\item [litc]( --- c ) $c\leftarrow mem.b[\mathrm{P}]$; $\mathrm{P}\leftarrow\mathrm{P}+1$\filbreak
\end{description}
<<address handling>>=
wire `L toaddr, incaddr, toR, R;
wire tos2r;
assign toaddr = inst[1] ? (inst[0] ? P : R) : A;
assign incaddr =
{ addr[l-1:1] + (incby | addr[0]),
~(incby | addr[0]) };
assign tos2r = inst == 5'b11100;
assign toR = state[2] ? incaddr :
(tos2r ? T : { P[15:1], c });
@
Memory access can't just be done word wise, but also byte wise. Therefore
two write lines exist. For byte wise store the lower byte of T is
copied to the higher one.
<<load/store>>=
5'b10000: begin
addr <= toaddr;
wr <= 2'b11;
end
5'b10100: begin
addr <= toaddr;
wr <= { ~toaddr[0], toaddr[0] };
T <= { T[7:0], T[7:0] };
end
5'b10???: begin
addr <= toaddr;
rd <= 1'b1;
end
@
Memory accesses need an extra cycle. Here the result of the memory
access is handled.
<<load-store>>=
if(show) begin
<<debug>>
end
state <= nextstate;
<<pointer increment>>
rd <= 1'b0;
wr <= 2'b0;
if(|state[1:0]) begin
<<store afterwork>>
end else begin
<<ifetch>>
end
<<next>>
@
There's a special case for the instruction fetch (the NEXT of the
machine): when the current instruction is a literal, we must use \texttt{inc\-addr}
instead of P.
<<next>>=
if(nextstate == 3'b100) begin
{ addr, rd } <= { &inst[1:0] ?
incaddr : P, 1'b1 };
end // if (nextstate == 3'b100)
@
<<debug>>=
$write("%b[%b] T=%b%x:%x[%x], ",
inst, state, c, T, N, sp);
$write("P=%x, I=%x, A=%x, R=%x[%x], res=%b%x\n",
P, I, A, R, rp, carry, res);
@
After the access is completed, the result for a load has to be pushed
on the stack, or into the instruction register; for stores, the TOS
is to be dropped.
<<store afterwork>>=
if(rd)
if(incby)
{ sp, T, N } <= { spdec, data, T };
else
{ sp, T, N } <= { spdec, 8'h00,
addr[0] ? data[7:0] : data[l-1:8], T };
if(|wr)
`DROP;
incby <= 1'b1;
@
Furthermore, the incremented address may go back to the pointer.
<<pointer increment>>=
casez({ state[1:0], inst[1:0] })
4'b00??: P <= !intreq ? incaddr : addr;
4'b1?0?: A <= incaddr;
// 4'b1?10: R <= incaddr;
4'b??11: P <= incaddr;
endcase // casez({ state[1:0], inst[1:0] })
@
To shortcut a \texttt{nop} in the first instruction, there's some
special logic. That's the second part of NEXT.
<<ifetch>>=
intack <= intreq;
if(intreq)
I <= { 8'h81, intvec }; // call $200+intvec*2
else
I <= data;
if(!intreq & !data[15]) state[1:0] <= 2'b01;
@
Here, we also handle interrupts. Interrupts are accepted at instruction
fetch. Instead of incrementing P, we load a call to the interrupt
vector (addresses from \$200) into the instruction register. The interrupt
routine just has to save A (if needed), and has to balance the stack
on return. Since three instructions can be executed without interrupt,
there's no interrupt disable flag internally, only an external interrupt
unit might do that. The last three instructions of such an interrupt
routine then would be \texttt{a! >a ret}.
\subsection{Stack Instructions}
Stack instructions change the stack pointer and move values into and
out of latches. With the 8 used stack operations, one notes that \texttt{swap}
is missing. Instead, there's \texttt{nip}. The reason is a possible
implementation option: it's possible to omit N, and fetch this value
directly out of the stack RAM. This consumes more time, but saves
space.\filbreak
Also, \noun{Chuck Moore} claims, that you don't need \texttt{swap}
--- if you don't have it, you help out with other stack operation,
and there's nothing to do, there's still \texttt{>a >r a r>}.
\begin{description}
\item [nip]( a b --- b )\filbreak
\item [drop]( a --- )\filbreak
\item [over]( a b --- a b a )\filbreak
\item [dup]( a --- a a )\filbreak
\item [>r]( a --- r:a )\filbreak
\item [>a]( a --- ) $\mathrm{A}\leftarrow a$\filbreak
\item [r>]( r:a --- a )\filbreak
\item [a]( --- a ) $a\leftarrow\mathrm{A}$\filbreak
\end{description}
<<stack operations>>=
5'b11000: { sp, N } <= { spinc, toN };
5'b11001: `DROP;
5'b11010: { sp, T, N } <= { spdec, N, T };
5'b11011: { sp, N } <= { spdec, T };
5'b11100: begin
rp <= rpdec; `DROP;
end // case: 5'b11100
5'b11101: begin
A <= T; `DROP;
end // case: 5'b11101
5'b11110: begin
{ sp, T, N } <= { spdec, R, T };
rp <= rpinc;
end // case: 5'b11110
5'b11111: { sp, T, N } <= { spdec, A, T };
@
If you don't want to live without \texttt{swap}, you can replace the
implementation of \texttt{nip} in the first line by:
<<swap>>=
5'b11000: { T, N } <= { N, T };
@
\section{Examples}
A few examples show, how to program this processor. Multiplication
works through the A register. There's one extra step necessary, since
each bit first has to be shifted into the carry register. Since \texttt{call}
clears carry, we don't have to do that here.\filbreak
<<mul>>=
: mul ( u1 u2 -- ud )
>A 0 #
*+ *+ *+ *+ *+ *+ *+ *+ *+
*+ *+ *+ *+ *+ *+ *+ *+
>r drop a r> ;
@
Division needs an extra step, too. Here, we need a real \texttt{swap},
but since there is none, we first use \texttt{over} and accept that
we have to use one extra stack item. Other than with \texttt{mul}
we here need to clear the carry after \texttt{com}. And finally, we
have to divide by two and shift in the carry.\filbreak
<<div>>=
: div ( ud udiv -- uqout umod )
com >r >r >a r> r> over 0 # +
/- /- /- /- /- /- /- /- /-
/- /- /- /- /- /- /- /-
nip nip a >r -cIF *+ r> ;
THEN 0 # + *+ $8000 # + r> ;
@
The next example is even more complicated, since I emulate a serial
interface. At 10MHz, each bit takes 87 clock cycles, to get a 115200
baud fast serial line. We add a second stop bit, to allow the other
side to resynchronize, when the next bit arrives.\filbreak
<<serial line>>=
: send-rest ( c -- c' ) *+
: wait-bit
1 # $FFF9 # BEGIN over + cUNTIL drop drop ;
: send-bit ( c -- c' )
nop \ delay at start
: send-bit-fast ( c -- c' )
$FFFE # >a dup 1 # and
IF drop $0001 # a@ or a!+ send-rest ;
THEN drop $FFFE # a@ and a!+ send-rest ;
: emit ( c -- ) \ 8N1, 115200 baud
>r 06 # send-bit r>
send-bit-fast send-bit send-bit send-bit
send-bit send-bit send-bit send-bit
drop send-bit-fast send-bit drop ;
@
Like in ColorForth, \texttt{;} is just an EXIT, and \texttt{:} is
used as label. If there's a call before \texttt{;}, this is converted
to a jump. This saves return stack entries, time, and code space.\filbreak
\section{The Rest of the Implementation}
First the implementation file with comment and modules.
<<b16.v>>=
/*
* b16 core: 16 bits,
* inspired by c18 core from Chuck Moore
*
<<inst-comment>>
*/
`define L [l-1:0]
`define DROP { sp, T, N } <= { spinc, N, toN }
`timescale 1ns / 1ns
<<ALU>>
<<Stack>>
<<cpu>>
@
<<inst-comment>>=
* Instruction set:
* 1, 5, 5, 5 bits
* 0 1 2 3 4 5 6 7
* 0: nop call jmp ret jz jnz jc jnc
* /3 exec goto ret gz gnz gc gnc
* 8: xor com and or + +c *+ /-
* 10: A!+ A@+ R@+ lit Ac!+ Ac@+ Rc@+ litc
* /1 A! A@ R@ lit Ac! Ac@ Rc@ litc
* 18: nip drop over dup >r >a r> a
@
\subsection{Top Level}
The CPU consists of several parts, which are all implemented in the
same Verilog module.\filbreak
<<cpu>>=
module cpu(clk, reset, addr, rd, wr, data, T,
intreq, intack, intvec);
<<port declarations>>
<<register declarations>>
<<instruction selection>>
<<ALU instantiation>>
<<address handling>>
<<stack pushs>>
<<stack instantiation>>
<<state changes>>
always @(posedge clk or negedge reset)
<<register updates>>
endmodule // cpu
@
First, Verilog needs port declarations, so that it can now what's
input and output. The parameter are used to configure other word sizes
and stack depths.\filbreak
<<port declarations>>=
parameter show=0, l=16, sdep=3, rdep=3;
input clk, reset;
output `L addr;
output rd;
output [1:0] wr;
input `L data;
output `L T;
input intreq;
output intack;
input [7:0] intvec; // interrupt jump vector
@
The ALU is instantiated with the configured width, and the necessary
wires are declared
<<ALU instantiation>>=
wire `L res, toN;
wire carry, zero;
alu #(l) alu16(res, carry, zero,
T, N, c, inst[2:0]);
@
Since the stacks work in parallel, we have to calculated, when a value
is pushed onto the stack (thus \textbf{only} if something is stored
there).\filbreak
<<stack pushs>>=
reg dpush, rpush;
always @(clk or state or inst or rd)
begin
dpush <= 1'b0;
rpush <= 1'b0;
if(state[2]) begin
dpush <= |state[1:0] & rd;
rpush <= state[1] & (inst[1:0]==2'b10);
end else
casez(inst)
5'b00001: rpush <= 1'b1;
5'b11100: rpush <= 1'b1;
5'b11?1?: dpush <= 1'b1;
endcase // case(inst)
end
@
The stacks don't only consist of the two stack modules, but also need
an incremented and decremented stack pointer. The return stack even
allows to write the top of return stack even without changing the
return stack depth.\filbreak
<<stack instantiation>>=
wire [sdep-1:0] spdec, spinc;
wire [rdep-1:0] rpdec, rpinc;
stack #(sdep,l) dstack(clk, sp, spdec,
dpush, N, toN);
stack #(rdep,l) rstack(clk, rp, rpdec,
rpush, toR, R);
assign spdec = sp-{{(sdep-1){1'b0}}, 1'b1};
assign spinc = sp+{{(sdep-1){1'b0}}, 1'b1};
assign rpdec = rp+{(rdep){(~state[2] | tos2r)}};
assign rpinc = rp+{{(rdep-1){1'b0}}, 1'b1};
@
The basic core is the fully synchronous register update. Each register
needs a reset value, and depending on the state transition, the corresponding
assignments have to be coded. Most of that is from above, only the
instruction fetch and the assignment of the next value of \texttt{incby}
has to be done.\filbreak
<<register updates>>=
if(!reset) begin
<<resets>>
end else if(state[2]) begin
<<load-store>>
end else begin // if (state[2])
if(show) begin
<<debug>>
end
if(nextstate == 3'b100)
{ addr, rd } <= { P, 1'b1 };
state <= nextstate;
incby <= (inst[4:2] != 3'b101);
<<instructions>>
end // else: !if(reset)
@
As reset value, we initialize the CPU so that it is about to fetch
the next instruction from address 0. The stacks are all empty, the
registers contain all zeros.\filbreak
<<resets>>=
state <= 3'b011;
incby <= 1'b0;
P <= 16'h0000;
addr <= 16'h0000;
A <= 16'h0000;
T <= 16'h0000;
N <= 16'h0000;
I <= 16'h0000;
c <= 1'b0;
rd <= 1'b0;
wr <= 2'b00;
sp <= 0;
rp <= 0;
intack <= 0;
@
The transition to the next state (the NEXT within a bundle) is done
separately. That's necessary, since the assignments of the other variables
are not just dependent on the current state, but partially also on
the next state (e.g. when to fetch the next instruction word).\filbreak
<<state changes>>=
reg [2:0] nextstate;
always @(inst or state)
if(state[2]) begin
<<rw-nextstate>>
end else begin
casez(inst)
<<inst-nextstate>>
endcase // casez(inst[0:2])
end // else: !if(state[2]) end
@
<<rw-nextstate>>=
nextstate <= state[1:0] + { 2'b0, |state[1:0] };
@
<<inst-nextstate>>=
5'b00000: nextstate <= state[1:0] + 3'b001;
5'b00???: nextstate <= 3'b100;
5'b10???: nextstate <= { 1'b1, state[1:0] };
5'b?????: nextstate <= state[1:0] + 3'b001;
@
\subsection{ALU}
The ALU just computes the sum with possible carry-ins, the logical
operations, and a zero flag. It would be possible to share common
resources (the XORs of the full adder could also compute the XOR operation,
and the carry propagation logic could compute OR and AND), but this
optimization is left to the synthesis tool.\filbreak
<<ALU>>=
module alu(res, carry, zero, T, N, c, inst);
<<ALU ports>>
wire `L sum, logic;
wire cout;
assign { cout, sum } =
T + N + ((c | andor) & selr);
assign logic = andor ?
(selr ? (T | N) : (T & N)) :
T ^ N;
assign { carry, res } =
prop ? { cout, sum } : { c, logic };
assign zero = ~|T;
endmodule // alu
@
The ALU has ports T and N, carry in, and the lowest 3 bits of the
instruction as input, a result, carry out, and test for zero as output.\filbreak
<<ALU ports>>=
parameter l=16;
input `L T, N;
input c;
input [2:0] inst;
output `L res;
output carry, zero;
wire prop, andor, selr;
assign #1 { prop, andor, selr } = inst;
@
\subsection{Stacks}
The stacks are modeled as block RAM in the FPGA. Therefore, they should
have only one port, since these block RAMs are available even in small
FPGAs. In an ASIC, this sort of stack is implemented with latches.
Here it's possible to separate read and write port (also for FPGAs
that support dual-ported RAM), and save the multiplexer for \texttt{spset}.\filbreak
<<Stack>>=
module stack(clk, sp, spdec, push, in, out);
parameter dep=3, l=16;
input clk, push;
input [dep-1:0] sp, spdec;
input `L in;
output `L out;
reg `L stackmem[0:(1@<<dep)-1];
wire [dep-1:0] spset;
`ifdef BEH_STACK
always @(clk or push or spset or in)
if(push & ~clk) stackmem[spset] <= #1 in;
assign spset = push ? spdec : sp;
assign #1 out = stackmem[spset];
`else
stackram stram(in, push, spdec, sp, ~clk, out);
`endif
endmodule // stack
@
\subsection{Further Possible Optimizations}
It would be possible to overlap memory accesses and operations on
the stack, since there are separate pointer registers. The understandability
of the code would suffer, and the critical path would also be somewhat
longer. With a guaranteed speed increase of 25\% (the cycle to fetch
instructions would vanish), and a maximum acceleration by 100\% (for
memory-intensive applications), this could be worth the trouble ---
when there's enough space.
If there's lack of space, it is possible to implement most registers
as latches. Only T needs to be a real flip-flop. For FPGAs, this is
not an option, flip-flops are cheaper there.
\subsection{Scaling Issues}
Two approaches allow to adopt the b16 to own preferences: word width
and stack depth. The stack depth is easier. The chosen depth of 8
is sufficient for the boot loader, but could cause problems for more
complex applications. Simpler applications however should fit with
a smaller stack.
The word width can be adopted for the application, too. A version
reduced to 12 bit (and also with a modified instruction set) is used
in a project at my employer Mikron AG. This required to change the
decoding of the instructions within the slot, and adopt the logic
to step over the first \texttt{nop}.
Furthermore, you can replace individual instructions. For the 12 bit
version, it was found that bit operations occur very frequently, and
byte accesses are completely irrelevant.\filbreak
\section{Development Environment}
I could present a longer listing here, this time in Forth. However,
I'll just describe the functions. All three programs are put into
one file, and allow interactive use of simulator and target.\filbreak
\subsection{Assembler}
The assembler resembles a bit \noun{Chuck Moore}'s ColorForth. There
are no colors, just normal punctation, as common in forth. The assembler
after all is coded in Forth, and therefore expects Forth tokens.\filbreak
Labels are defined with \texttt{:} and \texttt{|}. The first one automatically
call on reference, but can be put on stack with '. The last one more
resemble an interactive \texttt{Create}. Labels are only resolved
backwards. Literals must be taken from the stack explicitely with
\texttt{\#} or \texttt{\#c}. The assembler takes care of the ordering
within the slots. A \texttt{ret} is normally compiled with a \texttt{;},
preceeding calls are converted to a \texttt{jmp}. You can define macros
(\texttt{macro:} \ldots{} \texttt{end-macro}).
Also the well-known control structures from Forth can be used (must
be used for forward branches). \texttt{IF} becomes a \texttt{jz},
\texttt{jnz} is reached with \texttt{-IF}. \texttt{cIF} and \texttt{-cIF}
correspond \texttt{jnc} and \texttt{jc}. Similar prefixes are available
for \texttt{WHILE} and \texttt{UNTIL}.\filbreak
\subsection{Downloader}
A piece of block RAM in the FPGA is occupied by a small program, the
boot loader. This small program drives the LEDs, and waits for commands
from the serial line (115.2KB-aud, 8N1, no handshake). There are three
commands, starting with ASCII signs:
\begin{description}
\item [0]\emph{addr, len, <len$*$data>:} Programs memory from \emph{addr}
with \emph{len} data bytes
\item [1]\emph{addr, len:} Reads back \emph{len} bytes from memory starting
at \emph{addr}
\item [2]\emph{addr:} Execute the word at \emph{addr}
\end{description}
These three commands are sufficient to program the b16 interactively.
On the host side, a few instructions are sufficient, too:
\begin{description}
\item [comp]Compile to the end of line, and send the result to the evaluation
board
\item [eval]Compile to the end of line, send the result to the evaluation
board, call the code, and set the RAM pointer of the assembler back
to the original value
\item [sim]Same as \texttt{eval}, but execute the result with the simulator
instead of using the evaluation board
\item [check]( addr u --- ) Reads a memory block from the evaluation board,
and display it with \texttt{dump}
\end{description}
\section{Outlook}
More material is available from my home page \cite{web}. All sources
are available under GPL. Data for producing a board is available,
too. \noun{Hans Eckes} might make one for you, if you pay for it.
And if someone wants to use the b16 commercially, talk to me.
\begin{thebibliography}{1}
\bibitem{c18}\emph{c18 ColorForth Compiler,} \noun{Chuck Moore}, $17^{\mathrm{th}}$
EuroForth Conference Proceedings, 2001
\bibitem{web}\emph{b16 Processor,} \noun{Bernd Paysan}, Internet Home page, http://www.jwdt.com/~paysan/b16.html \url{http://www.jwdt.com/~paysan/b16.html}\end{thebibliography}
\end{document}