So I'm sitting at my desk, mouse in hand, digging through the guts of my Mac, trying to
track down yet another pathetic bug. The only trouble is, this is getting dull. I've done
it a thousand times, and I always win; it's just a question of how much time the old ball
and chain is going to eat up this time. Worse, the wind is blowing 20 knots right
outside my window, and for the windsurfers in the crowd, you know the exquisite
torture of good wind that you aren't allowed to transform into mind-blowing speed.
Maybe for you it's that you'd like to get home to see your insanely great mate. Or
you've got kids whose names you can't pronounce. It's even likely that some of you just
graduated from college and are still astonished that they pay you for something that's
so much fun. But you're thinking, "If I can get this bug fixed quickly, I can get back to
writing that rad Marathon hack to make the other net players slower than me."
Perhaps your boss has started to notice that you spend a lot of time on the job but you
don't really get very much done. Getting a little nervous? What if he's thinking of
pulling the plug on your baby because you're too slow?
Or maybe you shipped the 1.0 version, but it had a few too many bugs, andMacWEEK
was so incensed that they broke with the tradition of objective journalism and are
calling for your head. People who bought your software with actual money have been
calling every day, filling your answering machine with unveiled threats. It seems that
you left a bug in there that just cost several thousand people each about two weeks of
valuable time, reentering their data.
In particular, recognize that the wasted time from programming errors and bugs gets
exponentially more expensive the further into the process you get. If I make a syntax
error, I just fix it and recompile. If I toss buggy code over the fence to the testers, now
I'm wasting their time. If I ship software with occasional crashing bugs that I just
can't quite track down, I'm wasting thousands of people's time.
The point is, there are lots of ways to waste time while programming. I'm here today to
offer some ideas on how to save time through better programming habits, so that you
can take up windsurfing, or maybe the electric guitar, or learn how to pronounce your
kids' names, or gain "the power to crush the other kids" while playing Marathon.
Whenever I mention windsurfing, substitute your favorite quality-of-life enhancer.
I'll use some real life examples, and we'll see what sorts of lessons we can learn from
them. I've categorized these ideas in three ways. First, there are some obvious time
wasters that can be eradicated; these aren't really bug related, just daily time wasters.
Considering how much time bugs cost, the second category consists of high-value rules
that can find bugs quickly and painlessly. Finally, there are the super-value rules that
prevent bugs from happening at all. Be sure to consider how these ideas might apply in
your specific circumstances.
OK, I'm in the midst of writing the MMU tables for Blat, and I realize that I've already
got a similar table. Not being a fool, I know not to rewrite code I've already written, so
I open that file and casually copy and paste the table into my current work. Oh no, not
another copy and paste casualty! Apparently I missed changing those two table entries,
and with MMU tables that means the machine hangs before MacsBug loads. Seems like
every time I paste in code there's something I miss, making it not compile -- or
worse, causing a malfunction.
Even with small chunks of code, I've found it helps to review the pasted code line by
line, carefully, and not to assume that since it ran before, it'll run now. Some subtle
assumptions may have changed, and even though reusing code is certainly superior to
writing it again, don't be misled into thinking this is risk free. A more powerful
technique is to seriously modularize code, so that when I copy and paste I take an entire
routine, not just a few lines. With a well-defined interface, the chances of blowing it
are greatly reduced. This means adopting the habit of writing each routine with the
idea that I'm going to reuse it later. This radically improves every routine I write.
New rule: Reuse code modules, not code fragments. If the code has to be altered, inspect
it as if it were new (which it is).
While in the guts of Font/DA Mover, I ran across some very strange code that didn't
make any sense to me at all -- and it wasn't documented. It was never executed as far
as I could determine, but I painstakingly figured out that it was looking for System file
3.2 and, if found, would patch the OS to fix a font bug. I would have saved a full day of
effort had there been a comment in that funny little splat of code. Like I'm supposed to
know what the bugs in System 3.2 are off the top of my head?
Comments really are necessary to make code reusable and maintainable. I always
write "strategy" comments, which say what the routine is trying to do, and avoid
writing "tactical" comments, like what it's doing line by line. Remember, sometimes
the time savings occur in the future, not at the moment. I've found that skipping
comments is being penny wise and pound foolish. Usually the strategy comments help
clarify my thinking on the routine as well, so there actually is a short-term gain.
New rule: Always write strategy comments. It's possible to decipher intent from the
code, but why not just explicitly say it?
I used to think it was important to save every line of assembly code that was possible.
The first program I wrote for the Mac was Anaclock, an analog clock program, and I
remember thinking that if I changed the order of some routines I could save code. Don't
we all get into that mode sometimes? If I just change these two lines, I can save an
assignment, and blah blah. It must come from the old 128K Macs and Apple IIs.
Well, guess what? These machines are so stuffed full of junk nowadays that saving
just one or two lines is as meaningless an effort as trying to decide how many demons
fit on the head of a transistor. Worse, I spent my own valuable time deciding something
that has zero impact. Sorry, no can do anymore. My philosophy now is: write it
straightforward, easy to read, vanilla. I want to save my windsurfing time, not pretend
that I know up front what needs optimizing. In the Anaclock example, the computer had
an entire second between screen updates. When I actually measured execution time, all
the time was spent in CopyBits updating the screen, and waiting in the main event loop
for the next second to arrive. There was zero measurable time in my entire clock
calculation and offscreen drawing code.
New rule: No premature optimization. Measure with performance tools first. Then
optimize only where it counts.
During System 7 development, we once tracked down a bug, taking seven hours in the
middle of the night to find it, and it wound up being a bad parameter passed to a ROM
routine. Incredibly, I could have found that bug in about 15 seconds if I'd used the
Discipline tool. Nowadays, I never debug something by hand unless it has passed all the
debugging tools that Fred Huxham and I talked about in our article in develop Issue 8.
There are lots and lots of tools available now, and I use all of them. I don't care how
hard they are to use; if they can find a bug in seconds that might take me hours or days,
then I win. This includes such notorious tools as Blat and Jasik's debugger. I know
Blat's a pain, since it doesn't work on all machines, but hey, it's too valuable to skip.
Same with Jasik's debugger. Sure it's confusing, but it's got features no one else
provides. Before throwing the software over to the testers, I make sure it passes all
the tools.
High-value rule: Use the best tools, all the time. Don't spend time in a debugger when a
test tool will hand you the answer on a silver platter.
As part of a contract, my job was to make a program to save, print, and display 300
dpi bitmaps that were scanned in from a fax machine through new hardware. This was
to be a low-cost scanner, and my software would be the initial scan-and-display code.
Nothing too fancy, but it still required basic functionality. I bid 15 hours for the
entire program. Was I crazy? Well, of course, but not for this reason. I used MacApp
to give me the application functionality, and the FracApp300 sample program was a
good starting point for 300 dpi bitmap handling. All I really did was add an object to
talk to the scanning hardware, and I came in under bid!
Sometimes learning those new tough coding tools can really pay off. I generally try to
sample every new tool and coding advance that comes along to see if it can help me save
time. MacApp was clearly a massive win, because it focused my programming onto
teensy parts to be added instead of all the Toolbox calls of a typical application. In
addition, it was fully debugged and very robust, giving me a more solid final
application. I try not to be wedded to any given style or approach; I just want to use the
best stuff currently available.
High-value rule: Try new things. New ideas, approaches, tools, and programming
styles can be like winning the free-time lottery.
Sometimes it takes a while to recognize bad habits for what they are. While writing
Bowser, which turned into Mouser and then MacBrowser, I wrote the source code
parser by hand, to look for keywords. This was not a good strategy. It was quick and
dirty, and stayed dirty, and was less quick all the time. It would be reasonable to
expect that after modifying the parser for the eighth or ninth time to handle some
stupid language exception, I would have gotten a clue that this was not the right
approach. The right answer was to learn how the lex and yacc tools worked, since
parsers for both Object Pascal and C++ already existed in that format.
After seeing similar bugs go by several times, it becomes clear that something must
be done to stop that kind of bug. I don't want to spend time fixing the same problem
over and over again, so now my goal is to permanently fix bugs so that they can't
happen again. By this I mean changing how I do things, so that that specific bug will
either be caught quickly or never happen again. It can be as simple as adding a test to a
test suite to ensure that bugs of that form are caught immediately, or adding an assert
to catch that error. Or it can be as hard as changing my programming habits to never
use pointer math. Whatever it takes, I try to learn from each bug and make sure it
can't happen again. Especially after I've done something twice, it's time to write a tool
to fix that problem.
High-value rule: Learn from mistakes. If my dog gets bonked on the nose every time he
gets near the door, he learns to avoid the door. I want to be at least as smart as my dog.
Another slant on the Bowser problem is that I wasn't really trying to make the parser
right. If I'd been a little more quality conscious, I wouldn't have gone that route,
because it was clear that the hand-built parser was clunky. As noted before, the longer
a bug survives, the more expensive it will be. Early bug extinction is my goal, so I
consciously try to write with quality in mind. Examples are: using the strictest coding
rules, not using any tricky features of the compiler, using type-suggestive variable
names, insisting on type checking, not using raw pointer variables, avoiding type
coercion, adopting a simple easy-to-read style, writing clear module interfaces, and
using full warnings in the compiler.
Since I started noticing how much time bugs cost, I've changed my mindset on them. I
no longer automatically accept that code will just have bugs. I hate 'em. I want to kill
'em. Better, I want to kill 'em before they hatch. Since they take up my personal time,
I feel it's only proper to take it personally when they show up.
Super-value rule: Write with quality in mind. As they say, the inner game of
programming is so important.
Once upon a time, I was asked to fix a couple of bugs in Font/DA Mover and make it
work with TrueType fonts, as an interim solution before System 7. The program was
so disgusting to me that I just had to go in and clean it up. Move this here, change these
names, document some pieces, take out the redundant code, modularize some pieces --
ah, how aesthetically pleasing. Oops . . . I just introduced a couple of bugs while I was
"improving" the code. It felt like progress, but actually it was just motion. You know,
like company reorgs.
What to do? Don't "improve" code, unless it's never been debugged. Any fully debugged
code, no matter how shoddily written, is superior to newly written code, no matter
how pristine. It went against my grain, but the right answer was to leave it gross. That
heavily used Font/DA Mover code had thousands of hours of value in it, with literally
millions of testers, that were all lost when I rewrote it. Rewriting it took time that I
wanted to spend on something more valuable, like fixing the last few remaining bugs
-- and then getting outside and windsurfing! Once I rewrote the code, it was like a new
program, and thus needed a full development/testing/debugging cycle. I backed off to an
earlier "skanky" version and just debugged that.
Super-value rule: Never rewrite something that's been fully tested. It may be ugly,
but evolution is on its side.
We all know about the "cool" things that C can do, and some tricky ways of using it, but
sometimes isn't it a bit like juggling live weasels? When I found that using a #define
had added an extra unwanted character to each place I used it, it no longer seemed so
clever, and felt more like I was playing tricks on myself. Or how about that favorite of
putting an actual assignment in an if statement? It's cleverly camouflaged, but there
aren't any natural predators here, so I'm not sure this is needed. These simple
examples obviously don't do justice to the possible tricks that we've all seen, but they
all cost time and rarely add value.
OK, so it's clear that being "clever" often winds up being a way to play tricks on
myself. Is there anything wrong with doing it simply, in a straightforward, vanilla
style? I know for sure I'll get it done sooner and, even better, the programmer who
has to maintain this code won't have to waste a bunch of time understanding mindless
tricks (remembering of course that that maintenance programmer might very well be
me, two years after I forgot what tricks I was playing). And let's just forget the
malarkey about it saving code. Is it really worth saving 10 whole bytes out of a 16 meg
machine, at the expense of wasting my time? I want to count cycles and bytes only in
places where it makes a measurable difference.
Super-value rule: Write vanilla code. Doing it simply, and the same way each time,
also makes it more likely to be correct.
Back in the deep dark Macintosh past, I wrote the driver for an external RAM disk
called DASCH. This high-speed serial link required some different debugging tactics
than I'd used previously, because I couldn't step through the code; it was time critical.
Any slight perturbation in speed would overrun and cause an error, but I still needed
to debug it. It was like a "look Mom, no hands" type of debugging. Code inspection is OK,
but I wanted to be sure it worked as I read it. Have you ever read a piece of code that
took a branch you didn't expect?
The answer, although I didn't use the name at the time, was to use asserts. These have
been talked about a fair amount, and you've probably used primitive asserts under the
name of DebugStr. Nowadays, the most powerful combination I've used is to hook
together asserts with a failure handler like MacApp's catch/fail mechanism. Asserts
make it easy to build a debug-only version that checks every stupid thing that can go
wrong and lets me know right up front during testing, but doesn't compile into the
final version. The catch/fail stuff makes it easy to handle every possible error in a
graceful way. (See the article "Using C++ Exceptions in C" in this issue.)
If something absolutely positively cannot fail, I use a debugging-version assert to
catch the occasional times when it does fail, so that I can surprise myself early and not
spend hours tracking down the "impossible" error. One great thing to check with
asserts is input parameters, to catch those inevitable times when some routine passes
in rubbish.
Super-value rule: Use asserts along with a failure handler. Catching bugs as they
happen is vastly superior to backtracking 15 miles after the program crashes.
I'm not going to pretend that this is all there is to the idea of saving time, but hopefully
the idea seems worth pursuing. It has certainly helped me get better at my carving
jibes and, not incidentally, better at programming at the same time. Higher-quality
code, fewer bugs, earlier ship dates, happier customers, and more free time. Yup, I'd
say it's been worth it. If you've got some additional time-saving ideas, I'd naturally be
interested in trying them too, so write me at bo3b@rahul.net.
RECOMMENDED READING
BO3B JOHNSON (bo3b@rahul.net) is completely whacked out about windsurfing,
and takes summers off in order to windsurf every day. But since it's winter, he's doing
consulting so that he can pay for his next windsurf board and windsurf trip to Aruba.
Bo3b prefers to be addressed as "Bob," since the 3 is silent.*
Where's Dave? That other Johnson, who usually writes this column, is probably at the
public library researching his obsession du jour, taking his dogs for very long walks,
or reclining on the couch reading a book. Since he's cut back his working hours, we're
having guest Neophytes write this column. We can't promise they'll all be Johnsons,
however.*
Thanks to Jeff Barbose, Jim Friedlander, Brian Hamlin, Fred Huxham, Dave Johnson,
Jim Reekes, and Patty Walters for their terribly helpful review comments.*