View detail for Using the AT89C2051 MCU as Virtual Machine

View detail for Using the AT89C2051 MCU as Virtual Machine
Using the AT89C2051 Microcontroller
as a Virtual Machine
It is often cited that what differentiates an
embedded microcontroller from other
general purpose computing devices is its
integration into a larger electrical or electro-mechanical system. While this is
generally true, the fact remains that processors of widely differing capability and
architecture are employed in this regard.
Unfortunately, this broad explanation
defines nothing; we are still left to contend with everything from full-blown
embedded PCs to the smallest self-contained single-chip microcontrollers.
Within this expansive realm, conventional wisdom may lead to the conclusion that the smallest microcontrollers
are only appropriate for driving smallscale applications with very limited processing requirements. While this is
unquestionably the case in many
instances, a class of applications exists
that mandates a relatively high level of
program complexity within severely constrained space limitations. Faced with
such a seeming paradox, engineers
often feel they have no choice but to
adopt a less than optimal design strategy using a larger microcontroller than
originally intended.
The problem, of course, is one of limited
resources. Functional complexity implies
a non-trivial program, and the greater
the functional complexity the larger the
program. Even as the capability of small
single-chip microcontrollers continuous ly in ches u pwar ds, appl ica tion
requirements seem to grow at a commensurate rate. Trying to hit such a
moving target is difficult at best.
The economy of using a microcontroller
with just enough processing power for a
given application is a potent incentive to
find just the right fit. Of course, this only
works when the system requirements
are thoroughly understood and clearly
defined. Since such a design normally
has little reserve capacity, it is usually
hard pressed to handle features beyond
those originally specified. Should additional capabilities eventually become a
necessity, the result could be a system
that runs out of steam and an engineer
that runs out of options. Such are the
perils of designing on the edge.
Atmel’s AT89C2051 offers capabilities
that far exceed those of competing
devices of similar size. This opens up
potential design opportunities that were
simply unattainable with previously available parts. Housed in a 20-pin package,
Atmel’s miniature microcontroller retains
all the major features of the 8051 architecture. Furthermore, the AT89C2051
includes all of the 8051’s “special” pins
including the external interrupts, UART
transmit and receive lines, and the external timer controls. Even though the
AT89C2051 significantly ups the processing ante, it would seem that there
are limits to what you can accomplish
with any single-chip microcontroller.
This dilemma is nothing new. The traditional way of dealing with such limitations has been to operate the microcontroller in external memory mode. Common sense would indicate the hopelessness of applying such an approach to
the AT89C2051. After all, the
AT89C2051 is truly a single-chip design
that does not even possess an external
bus structure. It turns out that the situation is not hopeless at all.
AT89C2051
Flash
Microcontroller
Application
Note
w
5-47
Processor Simulation
Efficiency and Overhead
The concept of microprocessor simulation is widely used
and well understood. Simulation is often used for development
purposes where a PC program models a specific processor’s architecture and interprets and executes its binary
instruction set. Using this technique enables one to
develop, test, and debug algorithms that will ultimately be
combined into a larger program. Such a program will eventually run on a standalone microprocessor or microcontroller. Using simulation early in the design cycle is attractive
because it allows you to start developing code long before
the actual target hardware is available.
Processor simulation has also been applied to simulate
entire computing systems. In this context, existing application programs, in their native binary format, have been
coerced to run on various computers powered by completely different processors. For obvious reasons, the performance resulting from such an approach often proves to
be disappointing. This does not necessarily have to be the
case if the implementation is designed for a specific purpose. Factors effecting performance efficiency include the
host processor’s strengths and limitations, the specific
types of operations that are to be simulated, and, to an
extent, the language the original program is written in.
To illustrate that efficiency is a subjective matter, consider
what happens when a typical C program is compiled to run
on an 8051 processor. It’s inconceivable that, on such an
architecture, any C statement will effectively compile down
to any corresponding 8051 instruction. A single C statement invariably results in the execution of multiple instruction steps. It follows that, given an efficient simulated
instruction set, the simulation overhead might account for a
very small percentage of the overall execution time.
The key behind making this premise work is to devise an
instruction set and processor architecture that’s conducive
to performing the types of operations that a C compiler naturally generates. In such an implementation, the contrived
instruction set essentially amounts to an intermediate language. The op codes merely serve as a vehicle for succinctly conveying the compiler’s directives to the target processor for execution.
The target processor, while performing the functions of a
simulator, interprets the intermediate instructions to perform the functions specified in the original high level language source statements. The resulting efficiency can be
quite tolerable since the bulk of the instructions would execute regardless of whether they were emitted directly by
the compiler or invoked by the simulation kernel.
It turns out the performance penalty of such an approach
is, to a great extent, dependent on the way the program
memory itself is implemented. Since the AT89C2051 has
no external bus structure it makes sense to use a serial bus
to access the program memory. Using I 2C for this purpose
provides the required flexibility along with reasonable
throughput.
Selecting I2C as a memory bus presents the potential of
choosing from a wide variety of EEPROM memory devices.
The most favorable configuration is Atmel’s AT24C64 that
offers 8K bytes of storage in an 8-pin package. Utilizing
extended 16-bit addressing, the AT24C64 provides linear
access to the entire internal memory array. And although a
lot of functionality can be crammed into a single chip, additional devices can easily be added in 8K increments to handle very complex applications. Up to eight AT24C64s can
simultaneously reside on the I2C bus providing a full 64K of
storage while using just two wires.
Of course, serial memory access does come at a cost. In
this case the expense comes in the form of access time. To
an extent, this is moderated by the fact that the AT24C64
can operate at a 400 kHz clock rate (standard I2C is specified at a maximum of 100 kHz). Remember however, that
I2C can exact a significant performance penalty because a
substantial percentage of its bandwidth can be consumed
for control functions.
Virtual Processor Simulation
Many developmental simulators have been produced that
emulate the functions of popular processors and microcontrollers using standard desktop computers. The same principles can be utilized at the other end of the spectrum;
there are cases where running a simulation on a small
microcontroller can be put to an advantage. In this case,
however, the benefit is not derived from simulating a known
processor, but one that offers inherent advantages tailored
to solving the specific problem at hand. The implication, of
course, points to the design of a virtual processor. The idea
is based on the premise of using a real processor to implement a virtual device specifically designed to suit the special needs of a particular application. In other words,
designing the tool set for a particular job.
The fact is that adopting such a methodology can ultimately
result in an architecture that can be pressed to serve as an
efficient vehicle for a number of specialized tasks. Details
including the fundamental architecture, instruction set, and
memory model can be approached with total freedom. But,
can such an approach provide the level of performance
demanded by embedded applications?
5-48
Microcontroller
Microcontroller
The greatest overhead burden that I2C imposes involves
the transfer of addressing information. For every random
read or write, a 16-bit address must be transmitted along
with the extra overhead necessary to coordinate bus control for both the addressing phase and the data manipulation phase. Under such conditions, actual data movement
could be swamped by the requisite overhead resulting in
unacceptable performance degradation. Fortunately, I2C
provides a means of eliminating much of this wasteful activity.
The AT24C64, like all other I2C memory devices, contains
an internal auto-increment address generator. Using this
feature, once addressability is established, data can be
continually streamed in a sequential fashion. As each byte
is read and acknowledged the internal address generator
increments in preparation for the next byte transfer. The
AT24C64 sets the maximum speed limit at 400 kHz but I2C
does not impose a lower limit. Effectively, the minimum frequency can drop all the way to DC. As a result, it’s acceptable to suspend a sequential transfer for as long as necessary.
Utilizing these features, communications can be sped up
considerably. The ramifications are particularly significant
when the memory is used to store an executable program.
For example, once an address is written into the AT24C64,
data can be fetched in a continual stream until the program
branches or, if multiple AT24C64’s are used, until it
becomes necessary to cross into the next chip. At these
points it’s necessary to explicitly reload the internal address
generator. Normally, however, the majority of the accesses
will be sequential, resulting in greatly reduced overhead.
Processor Simulators and Language
Interpreters
It’s important to note the distinction between language specific interpreters that implement a defined language such
as BASIC, and a processor simulator that interprets a low
level binary instruction set. A tokenized BASIC interpreter,
while quite efficient in executing the commands that are
explicitly implemented as part of the language, is strictly
confined to what the language supports. The inherent efficiency of an interpreted language comes at the expense of
flexibility.
In contrast, a processor simulator, that deals with a true
binary instruction set, enjoys total freedom in combining
these basic op codes into larger functional entities in
almost limitless permutations. Just like a real processor, a
simulated processor can utilize its instruction set for standard and custom C library functions, floating point libraries,
device drivers, etc.
The Virtual Machine — An Imaginary
Processor
The processor to be described is imaginary in the sense
that its architecture and instruction set are original and
unique. Realize, however, that this is not just a toy or an
intellectual diversion—from an implementation standpoint it
is quite real. The fundamental concept has been successfully ported to a variety of processor architectures. A version exists that runs on a personal computer that is suitable
for demonstration and development purposes. The most
promising small-system port has been to the AT89C2051
due to the microcontroller’s standard processing core and
integrated peripheral set. The basic 8K Virtual Machine is
schematically depicted in Figure 1. The circuit’s simplicity
reveals that this is primarily a software implementation—
the definitive soft machine.
This imaginary processor, the product of Dunfield Development Systems, has served in various applications providing
reliable solutions to real world problems where a standard
configuration was not necessary, optimal, or practical. That
this Virtual Machine also goes by the name “C-FLEA”
affirms its optimization for efficiently rendering the output of
a C language code generator.
The prime currency of a processor is time. Viewed in this
context, the expense of complexity can prove unacceptably
burdensome. Taking this into consideration, the Virtual
Machine, based on a simple 16-bit architecture that incorporates only four registers, is the epitome of simplicity. This
register set comprises an accumulator, index register,
stack pointer, and program counter. Appendix A provides
detailed information about the Virtual Machine architecture
and instruction set. Refer to Table 1 for a description of the
fundamental resource set.
Although the Virtual Machine performs all operations to 16bit precision, the needs of many embedded systems
resolve to 8 bits. To facilitate working with this common
denominator, the Virtual Machine stores data in little endian
format (low byte first) which facilitates the use of a given
variable’s base address to refer to either an 8-bit or 16-bit
quantity. Interestingly, the architecture provides no user
accessible flags. When invoking a compare instruction,
internal flags persist only long enough to accommodate the
ensuing branch instruction or the intervening compare
modifiers (which are described later).
This spartan register set is made workable by the inclusion
of a variety of addressing modes that excel at the types of
stack manipulations that are central to the canonical C
implementation. The Virtual Machine’s memory access
instructions, detailed in Table 2, include the following
addressing modes: immediate (8 or 16 bit), direct, indirect
through index register (with or without offset), indirect
through stack with offset, top of stack, and indirect through
top of stack (remove or leave value on stack).
5-49
The bulk of the virtual instruction set is presented in Table
3. These instructions include memory access instructions,
arithmetic instruction, and logical instructions. In keeping
with the previously established proposition, most can return
either bytes or words.
Since the compare instructions are designed to only determine equality, the instruction set is augmented by a set of
special compare modifiers. Using these, nuances of relative (signed and unsigned) magnitude can be coerced from
the basic compare instructions. These modifiers are
described in Table 4.
Program branching is supported using the relatively conventional set of conditional and unconditional jump instructions shown in Table 5. Versions are provided for both near
and far destination targets to enhance code efficiency.
Note the inclusion of the SWITCH instruction which proves
especially useful since the “normal” compare instructions
destroy the contents of the accumulator when returning the
result of the compare operation.
Table 6 presents the stack manipulation set. Included are
common functions such as CALL, RETurn, and PUSH.
Conspicuously absent is an explicit POP instruction. The
corresponding functionality is provided by the various
addressing modes that, by default, manipulate the top of
the stack. For instance, POP A is synonymous with LD S+.
Additional instructions are included to facilitate stack frame
creation and destruction that is a necessary function of the
C language implementation.
Finally, the virtual instruction set is rounded with a number
of miscellaneous instructions shown in Table 7. For the
most part, these perform standard functions that should be
self explanatory. The input/output instructions are special in
that they offer an implementation specific avenue for establishing certain peripheral functions as instructions. Remember that, even though, the virtual instruction set offers the
programmer total freedom to construct any kind of computational sequence, all I/O operations are dependent on the
support coded into the Virtual Machine kernel. Essentially,
the simulation kernel is the software embodiment of a
microprocessor architecture. Naturally, the goal is to provide a general purpose engine capable of serving in a wide
variety of real embedded systems.
A significant number of op codes remain unassigned and
are available for future use.
5-50
Microcontroller
Initial Program Loader
While not actually part of the Virtual Machine, the simulation kernel contains a built-in program loader utility. This
operates serially and is invoked following a system reset by
a sequence of special commands from a utility program
running on the host computer. In addition to transferring the
load image to the Virtual Processor, the PC program provides a number of features which include a simulator (that
can hook into the target’s logical and physical I/O subsystem) and a console window for performing user I/O to
the target system. Since the Virtual Machine’s code generator emits a standard Intel HEX file format, the use of the
PC utility program is optional.
In principle, there is no reason why an AT24C64 cannot be
programmed externally using a standard device programmer just as you would program an EPROM for a use in a
typical embedded computer. Although workable, this
approach would, at the least, prove cumbersome throughout the development cycle. The difficulty of this approach
would be exacerbated in a system using multiple memory
chips. Obviously, it would be completely unworkable in the
event a Virtual Machine computer was rendered as a surface mount assembly.
Microcontroller
Figure 1. 8K Virtual Machine
+5
10 µF
1N914
10K
AT89C2051
+5
22 pF
1
4
14.7456 MHz
5
RST
VCC
20
.1 µF
XTAL2
XTAL1
22 pF
2
3
RXD/P3.0
TXD/P3.1
P1.7
P1.6
6
7
8
9
11
INT0/P3.2
P1.5
INT1/P3.3
P1.4
T0/P3.4
P1.3
T1/P3.5
P1.2
AIN1/P1.1
P3.7
AIN0/P1.0
10
19
18
17
16
15
14
13
12
GND
+5
AT24C64
VDD
1
2
3
8
A0
A1
NC
7
.1 µF
4.7K
4.7K
A2
SCL
SDA
4
+5
6
5
VSS
5-51
Virtual Machine I/O
The Virtual Machine handles physical I/O (as well as virtual
I/O) through the use of input/output instructions. It is natural
to reserve certain I/O addresses for on-chip functions such
as serial I/O and for access to the AT89C2051’s on-chip
parallel I/O ports. Additional I/O addresses are assigned to
second level functions such as serial port configuration and
direct I/O bit set and clear functions. The bit manipulation
functions are important when an on-chip parallel port is
simultaneously used for both input and output.
Consider the ramifications of performing a standard
read/modify/write operation on such a port. Normally this
would be accomplished by reading the port via an IN
instruction, performing a logical operation on the value, and
writing the modified data back to the port using OUT.
Should an input pin be externally pulling low while the port
was being read, the unfortunate outcome of this exercise
would be to render that line permanently jammed low and
unusable for any further input!
Additional virtual input/output devices are provided for functions such as time-of-day clock, general system timing,
pulse width modulation, and pulse accumulation. These are
implemented as background interrupt service routines and
are accessed as simple input/output devices.
Serviceable as the basic I/O resource set may be, it’s often
necessary to provide ancillary I/O functions external to the
processor. The Virtual Machine accomplishes this transparently by passing any undefined I/O addresses to the external peripheral trap. This handler uses a secondary I 2C bus
to implement an auxiliary external peripheral/memory channel.
Here, the instruction’s I/O address is taken as the I 2C slave
address. For output operations data is passed via the low
byte of the virtual accumulator. Input functions return data
in the low byte of the virtual accumulator. In both cases the
accumulator’s high byte is utilized to convey completion
status and can be interrogated to determine the outcome of
the requested operation. The result code reflects the status
of the data link transfer and either indicates valid completion or fault status. Should a fault be reported it could be
the result of a peripheral in busy status, a device that is not
present, or a legitimate peripheral malfunction.
Virtual Machine Assembly
To clarify the relationship between the Virtual Machine kernel, a virtual assembly language library function, and a virtual C application program, an example is in order. This will
also serve to illustrate how easily communication to the
outside world can be orchestrated in such an environment.
The program depicted in Listing 1 is a library function that
supports console I/O using a special I2C user I/O module.
(This is the same module that was detailed in the applica-
5-52
Microcontroller
tion note “A Framework for Peripheral Expansion.”) The
user I/O module contains a standard 20 x 4 LCD, 4 x 4 keypad, and beeper. These are supported using two I 2C-toparallel port expanders. The underlying premise is that,
once the data transport mechanism is hidden, the I2C ports
can be used just like any conventional I/O ports. In this
case the concealment is complete since the I 2C driver is
written in the AT89C2051’s native instruction set and is
therefore completely invisible and inaccessible to a virtual
program running on the Virtual Machine. Reading and writing to I2C devices now becomes strictly a matter of IN and
OUT.
Looking again to listing 1 reveals how virtual instructions
can be combined to generate a useful program. Far from
being constrictive, the virtual instruction set yields an economy of expression while retaining a great deal of flexibility.
The limited number of registers does, however, require a
reliance on the stack for parameter passing and for holding
intermediate results. This shouldn’t be surprising considering the fact that the Virtual Machine is primarily designed as
a C engine. Anyone familiar with the way a C compiler utilizes the stack frame should have little trouble adapting
these concepts to writing efficient assembler programs.
Virtual Machine Compilation
Not much can be said about the compilation process for the
Virtual Machine. This is truly a virtue since, after all, the primary purpose of a language compiler is to insulate the programmer from the complexities of a particular processor.
To those experienced with C compilers for 8051 processors, the most notable omission here is the absence of the
multiplicity of libraries for the various memory models that
are so necessary when working with a native 8051. Recall
that the Virtual Machine supports a single, eminently reasonable, flat 64K memory space.
Listing 2 reveals that there is nothing special and, more
importantly, that there are no artificial limitations inherent in
a C program written for the Virtual Machine. This program
implements a simple calculator function that uses the I2C
user I/O module as the system console device and utilizes
the long math functions from the Virtual Machine math
library. The actual functionality behind this module is secondary. What is more important is that it looks like a C program and behaves like a C program — and can be abused
like a C program. In short, it can be coerced to do the
things you need a typical embedded program to do.
Microcontroller
Pint Sized Computer
Although tiny by any scale of measure, the Virtual machine
behaves the way you would expect any self respecting processor to behave, virtual or not. More to the point, the Virtual Machine in actuality is a fully functional computer system. You would be hard pressed to find a smaller, fully
functional, computer with comparable capability that adequately supports the C programming language.
Using surface mount manufacturing techniques, a fully
operational computer can be constructed to fit into an area
the size of a postage stamp. The Virtual Machine’s large
program memory space, combined with its secondary I2C
memory/peripheral bus, makes the architecture suitable for
handling a number of relatively ambitious embedded
projects. Its minuscule size allows it to be placed anywhere.
Sources
If you are interested in experimenting with the Virtual
Machine concept, a fully operational PC based Virtual
Machine simulator, C compiler with libraries, and assembler are available for downloading from the the Dunfield
Development Systems bulletin board at (613) 256-6289.
For availability of the Virtual Machine processor, development system, and support software contact Mid-Tech Computing Devices USA; P.O. Box 218; Stafford, CT 06075
(860) 684-2442.
To obtain the listing 1 and listing 2 codes, please download
from Atmel’s Web Site or BBS.
5-53
Appendix A — Virtual Machine Architecture
Table 1. Fundamental Resource Set
ACC
16-bit accumulator 8-bit accesses are auto zero-filled
INDEX
16-bit addressing register, cannot be manipulated as 8 bits
SP
16-bit stack pointer
PC
16-bit program counter
Table 2. General Addressing Modes
Syntax
Coding
Description
#n
x0 ii(ii)
Immediate (8 or 16-bit operand)
aaaa
x1 dd dd
Direct memory address
I
x2
Indirect (through INDEX register) no offset
n,I
x3 oo
Indirect (through INDEX register) with 8-bit offset
n,S
x4 oo
Indirect (through SP) with 8-bit offset
S+
x5
On Top of Stack (remove)
[S+]
x6
Indirect through TOS (remove)
[S]
x7
Indirect through TOS (leave on stack)
Notes:
1. Addressing mode is in lower 3 bits of op code.
2. Mode S+ always pops 16 bits from stack. Only 16-bit values can be pushed.
3. Modes [S+] and [S] will always use a 16-bit address on the top pf the stack but the final target can be 8 or 16 bits.
5-54
Microcontroller
Microcontroller
Table 3. Memory Addressing Instructions
Name
Description
LD
Load ACC 16 bits
LDB
Load ACC 8 bits
ADD
Add 16 bits
ADDB
Add 8 bits
SUB
Subtract 16 bits
SUBB
Subtract 8 bits
MUL
Multiply by 16 bits
MULB
Multiply by 8 bits
DIV
Divide by 16 bits
DIVB
Divide by 8 bits
AND
And 16 bits
ANDB
And 8 bits
OR
OR 16 bits
ORB
OR 8 bits
XOR
XOR 16 bits
XORB
XOR 8 bits
CMP
Compare 16 bits (ACC = 1 if equal)
CMPB
Compare 8 bits
LDI
Load INDEX (16 bits only)
LEAI
Load INDEX with address
(00, 05)
ST
Store ACC 16 bits
(00, 05)
STB
Store ACC 8 bits
(00, 05)
STI
Store INDEX (16 bits only)
(00, 05)
SHR
Shift right (8-bit count only)
SHL
Shift left (8-bit count only)
Notes:
(Unused Address Modes)
1. ACC always contains 16 valid bits. All operations are performed in 16-bit precision. 8-bit operands are zero-filled when they
are fetched.
2. SI decrements when data is pushed
3. Data is stored in little endian format.
4. There are no user accessible flags. In the case of CMP, internal flags are maintaned only long enough to accommoate the
LT-UGE instruction.
5-55
Table 4. Compare Modifiers
Name
Description
LT
ACC = 1 if less than (signed)
LE
ACC = 1 if less than or equal (signed)
GT
ACC = 1 if greater than (signed)
GE
ACC = 1 if greater than of equal (signed)
ULT
ACC = 1 if lower than (unsigned)
ULE
ACC = 1 if lower than or same (unsigned)
UGT
ACC = 1 if higher than (unsigned)
UGE
ACC = 1 if higher than or same (unsigned)
Notes:
1. These instructions must immediately follow a CMP instruction.
2. NOT instruction is used to implement explicit NE.
Table 5. Jump Instructions
Name
Description
JMP
Long jump (16-bit absolute)
JZ
Long jump if ACC=0 (16-bit absolute)
JNZ
Long jump if ACC!=0 (16-bit absolute)
SJMP
Short jump (8-bit PC offset)
SJZ
Short jump if ACC=0 (8-bit PC offset)
SJNZ
Short jump if ACC!=0 (8-bit PC offset)
IJMP
Indirect jump (Address in ACC)
SWITCH
Jump through switch table (ACC=value, INDEX=table)
Note:
1. Switch table format: addr1, value1, addr2, value2, ... 0, default addr
Table 6. Stack Manipulation Instructions
Name
Description
CALL
Call subroutine (16-bit absolute address)
RET
Return from subroutine
ALLOC
Allocate space on stack (8-bit value)
FREE
Release space on stack (8-bit value)
PUSHA
Push ACC on stack
PUSHI
Push INDEX on stack
TAS
Copy ACC to SP
TSA
Copy SP to ACC
Note:
5-56
1. Explicit POP instruction are not required since various addressing modes use and remove the top item on stack.
Microcontroller
Microcontroller
Table 7. Miscellaneous Instructions
Name
Description
CLR
Zero ACC
COM
Complement ACC (ACC = ACC XOR FFFF)
NEG
Negate ACC (ACC = 0 - ACC)
NOT
ACC = 1 if ACC was 0, else ACC = 0
INC
Increment ACC
DEC
Decrement ACC
TAI
Copy ACC to INDEX
TIA
Copy INDEX to ACC
ADAI
Add ACC to INDEX
ALT
Get alternate result from MUL/DIV
OUT
Output byte in ACC to PORT
IN
Read byte from PORT
SYS
System interface function
Note:
1. ALT obtains the remainder after DIV and obtains the high word after a multiply. This instruction must be executed immediately after the MUL or DIV.
5-57
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement