MACHINE CODE WITHOUT TEARS

The Niche takes a first look at Basic Compilers for the Spectrum.

There's a world of difference between ZX BASIC and machine code. Programs in BASIC tend to be slow, with jerky graphics and poor sound effects. Programs in machine code run hundreds of times faster - which doesn't mean that the graphics just jerk around the screen at Warp Factor Ten. No, machine code programming permits smoother movement, '3D' perspectives, simultaneous sound, animation and so on.

Of course, machine code doesn't necessarily make a game playable, and some classic games have been programmed entirely in BASIC - Mined Out, Football Manager and Velnor's Lair for instance. But if you want to write a shoot-em-up game, or a program with sophisticated graphics, you'll almost certainly find BASIC too slow.

Back in the olden days of Spectrum programming, when programmers assembled code in their heads, and the back pages of the old orange manual were the first to fall out, there wasn't much alternative to learning machine code once you'd come up against the limitations of BASIC.

Learning machine code was a traumatic process: before disks and microdrives (which brought their own meaning to the term random access') every programming mistake meant a crash. Teaching yourself machine code was a frustrating process as, after each mistake, it took several minutes to re-load your assembler, debugger and program source from tape in preparation for another crack at the problem.

Predictably it wasn't long before someone figured out that, if the computer was so fiendishly clever, it really ought to be able to make up machine code for itself. In this article we take a look at currently available BASIC to machine code translators, or 'compilers'. Next month we hope to examine Colt and Blast, two new and aggressive-sounding BASIC compilers which are under development. In future Niches we'll blow the dust off other Spectrum languages such as Logo, Forth, C, and Pascal.

Meanwhile, back at the keyboard....

WHY IS BASIC SO SLOW?

BASIC is slow because everything you type is carefully checked to make sure it is correct. This would be fair enough if it only happened once, when the program was entered, but the exhaustive checking continues even while a program is running.

If you write a program in BASIC to add 2 and 2 a hundred times, the computer will take as long to work out the answer the last time as it did the first 99 times. The actual adding is done fairly quickly in machine code (which is the only language the Spectrum's Z80 processor can really understand), but the overall effect is still very slow. This is partly because BASIC checks the syntax of lines over and over again, even after they have been entered (in case some stray POKE or Cosmic Ray has changed the contents of program memory?)

BASIC is also hampered by the need to cope with all sorts of special cases. The routine to add numbers in the Spectrum ROM has to be able to cope with functions, arrays, numbers and variables; these can have almost any value from minus several zillion upwards. The Z80 processor can only cope with a few digits at a time; it has to do all its arithmetic in several steps, just in case. Worse still, it can't multiply and divide at all, so these operations must be performed 'longhand'.

Much of the code in the Spectrum's ROM is taken from the earlier ZX81 BASIC, which was squashed into just 8K. In order to keep the size down, parts of ZX BASIC were written using a leisurely version of the compact Forth language, rather than machine code. A new ROM for the Spectrum is planned (though not by Sinclair), but nothing has materialised yet.

One of the nice features of Spectrum BASIC is the way that it lets you type in new lines of program and scrub out old ones as you test your program. This is hell for the BASIC system, which has to keep scrabbling around in tables to keep track of shifting variables and program lines. The longer your BASIC program, the worse this gets, so that a 20K program may run at half the speed of a 2K one. In compiled BASIC, however, the position of every line and variable is fixed. This makes programs fast, but means that you have to re-compile the whole lot if you changed one line.

Finally, ZX BASIC is cursed by the stupid way humans like to write things. We write 'X = 6 + 7' when the computer would be much happier with '6 + 7 = X'. It can't do anything with the name X till it finds the equals sign (meaning that a value must be stored). Similarly, the equals sign isn't really relevant till the computer knows what is to be stored. The plus sign means add two values - there's no point telling the computer about it until it has found both the numbers. ZX BASIC actually performs calculations in the second sequence (which is called Reverse polish Notation), but it has to re-order them from the first sequence every time it finds them, and that is a slow process.

WHAT'S A COMPILER

The Spectrum BASIC compilers are programs which read a BASIC program and produce a machine code equivalent, The compiler and both programs have to be in memory all at once, which limits the size of compiled programs to 10-20K.

Compiled code may be anything from 2 to 200 times faster than the original, depending upon the compiler you are using and the intricacy of the original program. We ran (or at least, tried to run!) the standard BASIC benchmark programs on each of the compilers. The results are shown in the Timing Table, along with the published timings for ZX BASIC.

The timings are not as fast as for pure machine code, which allows much more freedom to the programmer, but they are easily fast enough for most games programming. A number of commercial games are written in compiled BASIC (including Frank 'N' Stein, published by PSS) and look none the worse for it, although you'd be hard-put to write Knight Lore with a compiler.

A few of the positions in the table contain asterisks, because the test program could not be processed by that compiler. In order to keep compiled programs fast, and reduce the complexity of the compiler, the packages all impose restrictions on what they can compile.

COMPILER CONSTRAINTS

Softek's FP compiler is the only one that can cope with decimal values, for instance - this makes it much slower than the others, but means that it is the only compiler suitable for use in business programming. But who wants to run as business on a Spectrum anyway? The other compilers restrict you to whole numbers between -32767 and 32767, although you can use values up to 65535 in POKEs and suchlike.

You can switch back and forth from normal BASIC, machine code and compiled code with USR calls and RETURN instructions, so it is possible to write programs in a mixture of languages if you need speed at one point and sophistication elsewhere.

The Softek compilers (FP and IS)are the only ones which allow you to use full BASIC string-handling; Mcoder gives you a fairly complete set of facilities to work with short strings (up to 255 characters) but Zip and the Mehmood compiler can only offer simple routines to read and write characters. You could probably write a text adventure using Mcoder or one of the Softek compilers, but you'd be much better off using The Quill.

Array handling is similarly limited - none of the compilers allow arrays of more than one dimension,and the IS and Mehmood programs won't allow arrays at all, you can use long variable names with Mcoder and the Softek compilers, but the cheaper packages restrict you to 52 short variable names.

The 'core' of ZX BASIC commands - PRINT, INPUT, PLOT, DRAW, LET, GO SUB, IF, and so on - are allowed by all the compilers. The Mehmood compiler doesn't allow FOR loops, which meant that we couldn't run some of the benchmark programs.

One of the snags of real machine code is the fact that you can't 'break in' to programs. This suits software houses, who want to discourage piracy, but it is very inconvenient for programmers. The only way you can stop a machine code program is to pull out the plug and re-load it. Zip and Mcoder allow you to break into compiled programs at will, but the Softek compilers require a special command wherever you might wish to break into compiled programs, You can't break into programs produced by the Mehmood complier at all.

The Softek compilers allow you to put special instructions in REM statements. These instructions only work once a program has been compiled, which is inconvenient since you can't test such programs fully in normal BASIC - one of the big advantages of BASIC compilers over machine code is the fact that you can test your programs interactively, with all the BASIC checks and hand-holding to help you, before you compile them.

Softek's special instructions allow you to check for the Break key, enter machine code into the program, and move simple (character-sized) sprites smoothly around the screen. On the FP compiler you can also trap errors and simulate the ON GO TO statement. None of the compilers let you GO TO a calculated line number - you must always GO TO a specific number.

Mcoder offers some REM instructions, but these are designed for program testing. You can turn off BREAK checks, giving marginally faster code, or turn on a 'trace' facility which shows the current line as it is executed. Mcoder and Zip allow you to pass variable values back and forth between BASIC and machine code.

THE COMPILERS COMPARED

So far we've taken a broad overview, looking at the compilers together. In the following section we look closely at each of the five Spectrum compilers (there were six, but the first Spectrum BASIC compiler, SUPER C is no longer available).

None of these compilers really offer 'instant translation' for your BASIC. With the possible exception of Softek's FP you really have to write your program with compilation in mind - it is hard work to convert existing BASIC to suit any of the compilers. Also, there are some things which are hard to do without the flexibility of real machine code. That said, the packages all produce working code pretty effortlessly, and you can be reasonably confident that compiled programs will work first time - unlike hand-coded ones!

Next month, PR companies willing, we should be able to report on two new compilers - Colt, from HiSoft, which is a development of Mcoder, and Oxford Computer Systems' Blast, which promises to compile absolutely any ZX BASIC program, without alteration. At the moment we're having a bit of trouble wheedling copies out of the manufacturers - they both seem to be holding back until they've had a chance to dismantle their competitor's product! We'll compile more information next Niche....

FP COMPILER
£19.95
Softek, 12/13 Henrietta Street, Covent Garden, London

This is the most expensive compiler by a clear tenner, so it had better be good - or at least different! As the benchmark timings show, it produces fairly pedestrian code, typical 2-10 times faster than normal BASIC - but it is very flexible. You can use FP to speed up almost an ZX BASIC program that doesn't use arrays of more than one dimension, or the VAL and VALPP functions. The compiler also disallows calculations in DATA and GO TOs, but we wouldn't dream of using those, would we?

The documentation is barely adequate - a single large sheet of paper with an introduction, list of compiled statements and brief technical discussion.

The FP compiler displays the current line being processed as it works. When an error is found the compiler stops and shows the line containing the problem, with a question mark to show where the problem was found. You can't go on to detect subsequent errors, but this doesn't matter much since the compiler is very fast. You can compile several programs into different areas of memory by using CLEAR between on compilation and the next.

FP is a well-written program, but it is expensive and may not be useful to many Spectrum users, since it doesn't offer a dramatic speed increase over well-written BASIC. We'll look at it again next month, when we examine BLAST, a new compiler designed to process 'off the shelf' programs.

COURTING COMPILERS

GO-FASTER!

FLIRTING WITH FORTH

Once the initial association with Sinclair Basic has run its course, Spectrum owners soon come to realise its limitations - especially with regard to speed of operation. Steve Mann checks out alternative routes - in particular the 'FP'and 'IS' compilers from Softek, and the Abersoft Forth package.

All Spectrum owners will have noticed that their implementation of, say, Missile Command just doesn't have the speed or the smooth, pixel-by-pixel movement that a true arcade game devotee has come to expect from a program. In fact, Basic is just about okay if the display is kept simple - PRINT AT is actually, relatively, quite fast if the number of objects that need to be moved is kept small - but once the number of moving objects reaches four or five, the movement slows down dramatically.

Sinclair's highly individualistic - some might say bloody-minded - method of laying out the display file ensures that direct POKEs to the screen are more trouble than they're worth, so something different is needed. Another drawback is that anything that is moved a character square at a time is going to look jerky when compared with the single pixel steps achieved in professional programs.

WHICH WAY OUT

A recent innovation is the idea of a 'games design' package; Melbourne House has just marketed one of these, as has Quicksilva. But these are necessarily limited - the Quicksilva package, for example, while allowing the designer a fair amount of freedom in manipulation of sprites, insists that the resultant game is of one of four types, described as Space Invader, Asteroids, Scramble or Berserk. If you want a completely new format then, tough, you're out on your own, buster.

Pundits will pontificate patronisingly... "Where's the problem?" they'll ask, "machine code will deal with all that and more" - but no one likes a smart-arse, and machine code has a reputation for sending beginners gibbering and whimpering back to the comfort of Basic. It's not that it's amazingly difficult - in fact, in some ways it's extremely simple - but it's not an easy subject to learn from books and early attempts at understanding often prove so off-putting that the whole idea gets quietly shelved in the 'forget it' tray.

There has to be an easier way, and there is - in fact there are several. One way out is to learn Forth - definitely the year's most fashionable language for home computer users - and later on in this article we'll be taking a look at the Abersoft Forth compiler for the Spectrum. Another alternative is to use compiled Basic instead of the built-in interpreted version and, as Softek has released new versions of its Basic compiler, it seems only right and proper to give these the once-over.

So, introductory waffle over, it's time now to move towards the highly laud- able aim of making life easy for the user. Let's get down to some specifics...

THE COMPILER ROUTE

Integer and Floating Point From Softek

Purists may look down on Basic compilers as being something of an easy way out, but then who wants to make things difficult? The principle behind a Basic compiler is perfectly straightforward - just type in your Basic program (or load it from tape or Microdrive), hit a couple of keys and, presto, the 707 has become a space shuttle - at least that's the theory.

Of course, nothing is actually that simple. The manual for the Softek compilers claims that compiled programs will run 'up to' 200 times faster (in some circumstances even 500 faster is possible), but the important point to note is that 'up to' is apostrophised (if there is such a word). In practice you will find that the compiled code is rarely that efficient and your programs are unlikely to be any more than 10 times faster. But let's not be ungrateful, a speed increase of 1000 percent is certainly not to be sneezed at!

The other point to note is that there is a trade-off of speed against ease of use - the Softek 'FP' floating-point compiler offers many more features to the user; it will handle almost all Basic commands and functions, including strings and arrays (single dimension only) and a full range of 'real' numbers. The 'IS' integer compiler has many more restrictions on its use - no arrays, fewer commands and integer numbers only, in the range - 32767 to 32767 (some functions will store a number in the range 0 to 65536 - for example, PEEK, POKE and USR). Floating point numbers are stored in five bytes, whereas integers take up two only - so the IS compiler produces much quicker code than does the FP version; in some cases operation can be speeded up by using the INT function, but generally the FP compiler will run two to ten times faster and in all but optimum conditions the speed increase will be at the lower end of this range.

In most instances, the 'IS' model is perfectly satisfactory, despite its more limited range of facilities, and the speed factor more than makes up for the extra care that is needed to ensure that the Basic is compilable. The lack of floating point maths need not necessarily be a drawback - integer numbers are often quite sufficient and, indeed, most Forth implementations will not handle floating point numbers anyway.

As a demonstration of the comparative speeds of interpreted and compiled Basic, let's examine a very simple program that plots every point on the screen (Figure 1). A couple of FOR/NEXT loops do the job quite adequately and, as can be seen from the timings, the 'IS' compiled version is fastest by many orders of magnitude. In the segment of this review that looks at Abersoft Forth, I have included a Forth version of this routine and a timing for further comparison.

MINUS MASTER KEY

Both Softek compilers use the same memory addresses, and both take up approximately 6K of RAM. I had hoped to incorporate Softek's Master Key program, which defines the SPACE key as a further SHIFT and which then uses this in conjunction with the numeric keys to give user-definable function keys. Unfortunately Master Key, which worked a treat on my earlier Spectrum model, is another fine program that has fallen by the wayside now that Sinclair has started mucking about with the ULA (see Front Lines, Sinclairwatch). Your author's much-loved Issue 2 Spectrum gave up the ghost a few weeks ago and despite having now got a replacement - also categorised as an Issue 2 - affections for the new model are already strained by the number of programs that now fail to work. Sinclair Research may claim that it's all the fault of software manufacturers who have jumped to conclusions about specifications, but one program that resolutely refuses to run on my new machine is Sinclair's own IQ Test! Anyway, Master Key would have been a boon when used in conjunction with the compiler. It would have enabled Basic programs to be compiled at a single keypress, thus obviating the need to remember the somewhat clumsy RAND USR command needed both to compile and then to run programs. I just hope that Softek produces a revised version for later Issue 2 and Issue 3 machines.

Anyway, even without Master Key, Softek's compilers work well enough. The compiler is first loaded and the user is asked if RAMTOP is okay at 40000 (48K version). RAMTOP may be moved to allow many different short routines to be linked together. After typing in the Basic, all that is needed is to key in RAND USR 59300. The compiler makes two passes through the program, printing on-screen the start and end addresses and detailing any errors. Compilation stops at any mistake and the offending line may be pulled down and edited in usual Basic fashion. Once any errors have been dealt with, all that's needed is to make a note of the start address and use this in another RAND USR call to RUN the program. The Basic program remains in memory if desired, so both interpreted and compiled Basic programs may be run.

Initially, all but the simplest programs will probably refuse to compile when using the 'IS' version - but it's not too difficult to rewrite the offending sections to allow for the limitations (arrays will obviously take some thought - one solution is to POKE the required information into spare memory for later retrieval). Only skimpy, provisional documentation was available to work with while writing this review, but Softek assures us that full manuals will be available very shortly. The 'FP' compiler is much less finicky, although it did seem impossible to use a third parameter in a DRAW statement, despite the manual's claims to the contrary.

One particularly nice feature is the use of special REM statements to give commands and facilities not normally available in Sinclair Basic. These include such features as ON ERROR GOTO; ON n GOTO, where n must be a single-letter variable: and, most useful of all, REM S a,x,y which plots the character whose ASCII code is at PLOT position x,y. This REM statement allows some very smooth pixel-by-pixel movement (sec Figure 2) and can be used with user-defined graphics to produce mini-sprites.

There were some slight problems with graphics when using the compilers - the Union Jack program from the Sinclair manual, for example, gained some extraneous black blobs when compiled - but there was certainly nothing that could not be put right after a little thought. Hopefully the final version of the manual will go into a little more detail than the preliminary effort - but even with just this, both Softek compilers proved easy to use and gave some impressive results.

It's also nice to see that Softek has dropped its demand for royalty payments on programs that are produced with its compilers - all the company now asks is that the use of a compiler is acknowledged on packages and advertising.

The 'FP' Compiler costs £19.95, the 'IS' Compiler £9.95 and both together the price is £24.95. You can get them from Softek, Suite 60, 12/13 Henrietta Street, London WC2.

SOFTEK 'IS' AND 'FP' COMPILERS

These are compilers for the 16K and 48K Spectrum - the 'IS' being an integer compiler, and the 'FP' a full floating point compiler. As these are similar in many ways, they will both be covered in this one review.

Each of the compilers contains 16K and 48K versions on the same tape and comes in a large video case. They are supplied with a short but informative instruction leaflet. While the packaging looks very game-like, any more than a cursory glance will show that they are in fact very comprehensive and useful utility programs.

To load the compiler you simply type LOAD "". The program loads in two parts, a BASIC loader then the compiler itself. When fully loaded you will be asked whether RAMTOP at 40000 is OK, and given the option to change this if necessary - your BASIC will be compiled immediately above RAMTOP.

Once RAMTOP is set, you are told that you can NEW and BASIC loader.

When the compiler is loaded, you can type in your BASIC program, or load it from tape or microdrive in the usual way. As you may expect, the 'FP' version will handle more of the BASIC than the 'IS', and your BASIC program will have to be written bearing in mind the list of allowed commands. Both compilers are quite comprehensive, in the string handling, for instance, strings are totally flexible and can be any length. String slicing is allowed as are LEN, STR$ CODE CHR$ INKEY$ SCREEN$ "+" (for concatenation). Strings can be compared as in BASIC, this is far more powerful than the string handling in the MCODER compiler. When compiling BASIC programs from magazines and so on then obviously some modification is usually needed, but both compilers handle most BASIC commands and it is usually quite easy to make the required changes to programs.

When your BASIC is loaded and ready to be compiled you simply type RAND USR 59300 (26600 for 16K).

Various messages appear as the compiler passes twice through your BASIC - these are:

1. START ADDRESS. This is always one byte above RAMTOP.

2. END ADDRESS.

3. VARIABLES END. This message appears when your program contains string or arrays. These are stored immediately after your compiled BASIC.

4. FIRST PASS/SECOND PASS.

5. ERROR/NO ERRORS.

After a successful compilation the message 'NO ERRORS' appears. If the compiler finds an error then the 'ERROR' message will appear, compilation will stop and the offending line is displayed with a question mark after the error. If this occurs then pressing EDIT will bring down the erroneous line for correction.

If compilation is successful your program can be run using RAND USR START ADDRESS (given in 1. above).

Error reports use standard Sinclair error codes.

Your BASIC program remains in memory after compilation. This allows modifications to be made if the machine code does not run as planned.

If all is well your compiled program can be saved as CODE in the normal way. Note, however, that the compiler must be saved along with the compiled code as it contains several runtime routines needed for efficient programs. As with the MCODER compiler, programs can be stacked in memory and used as subroutines; this is done by carefully setting and resetting RAMTOP.

SPECIAL FEATURES

REM statements are used in both SOFTEK compilers to obtain facilities not available from BASIC. These include:

'IS' Version
REM B - Checks to see if BREAK is pressed
REM S,a,x,y, - PRINTS a character (ASCII code 'a') on to the screen at PLOT poisition x, y.

'FP' Version
REM O,a,n,n - This simulates the BASIC function ON a GOTO n,n...
REM E,n - If an error occurs jump to line n.

IN USE

Both compilers proved very easy to use and the speed improvements is immediately noticeable. There seemed to be some problem when trying to use READ and DATA statements within a program loop - the program in question worked under BASIC but gave an OUT OF DATA message when compiled. Also if a variable was assigned within a FOR-NEXT loop an OUT OF MEMORY error was sometimes encountered during compilation. This seemed to be intermittent and if it did occur, using CLEAR cured it. These minor irritations aside, both compilers worked OK.

APPLICATIONS

Both compilers take up about 6K of RAM. The FP compiler is obviously capable of more accurate results in scientific and technical applications but also, as it uses time consuming (accurate) floating point notation, it is slower than the 'IS' version. The 'IS' compiler is more suitable for such things as games.

SOFTEK claim in their advertisement, 'Superfast games involving hi-res graphics and movement are now possible even if you are only knowledgeable in BASIC.' This is to some extent true, but the hi-res movement (using REM S) ranges from not particularly fast to downright slow. It is just not fast enough for fast-moving games. Both compilers would be ideal for use alongside other machine code subroutines with, for example, an interupt driven sprite-moving routine. This would obviously need some knowledge of machine code, so it wouldn't be quite lair to say that compilers are the complete alternative.

CONCLUSIONS

At £9.95 for the 'IS' compiler, and £19.95 for the 'FP' version these programs represent excellent value for money. If you are looking for a compiler these two are probably the best on the market at the moment. The 'IS' compiler is far superior to the MCODER from PSS, with the added bonus of (slow) mini-sprite graphics. Please note the copyright notice on the compilers, however - that the use of SOFTEK'S compiler must be clearly acknowledged on packaging, advertisements and on the initial screen title display if it is used for creation of commercial software.