Hacker News Reader

Comment by sillysaurusx | Hacker News Reader

I happen to agree, but only in this special case. Normally a giant switch statement is a pain to work through if there's even a tiny amount of logic involved.

One hilarious case was at Scottrade. I was poking around their C++ codebase one day and to my amazed horror saw a file hundreds of lines long, and a big chunk of it was devoted to handling leap years. Like, every leap year, one at a time.

There are few enough leap years that it should in theory be difficult to write hundreds of lines of code to handle each of them, but some contractor managed to.

And so the ball of mud continued. What can you do but laugh and embrace the horror?

View on HN

I think any time you're dealing with data in the form of an enum, a switch usually the natural way to handle cases. Many compilers will warn you when you've missed a case, which is a nice check to have. Similarly, in languages that support ADTs, pattern matching tends to be the sensible thing to do.

But I agree that in the case you described, that the programmer was being stupid. However they could have written it as a giant if-elseif block and that would also have been stupid, or a loop over a giant list of years, and that also would have been stupid. I think the problem was the programmer not thinking about the problem carefully, not with the control-flow construct they used to write the bad code.

> What can you do but laugh and embrace the horror?

laugh. embrace and share the fun with others: submit to https://thedailywtf.com

Maybe the author wanted to squeeze as many hours out of the task as they could? Plenty of non-top-quartile people love this sort of thing as a day at the beach.

For the diametrical opposite, here's one time on an embedded system where I really didn't want to bring in a library dependency, so I wrote an inconspicuous-looking good-enough version:

   y = (d - (d + 366 + (d >= 47847)) / 1461 + (d >= 47847)) / 365;
   d = d - (d + 365 + (d >= 47848)) / 1461 + (d >= 47848) - y * 365;
   y += 1970;

#include <iostream>

int main() { long long d; std::cout << "Enter a Julian Day Number: "; std::cin >> d;

    if (d < 0) {
        std::cerr << "Error: Invalid Julian Day Number (must be non-negative)." << std::endl;
        return 1; // Indicate an error to the system
    }

    const int DAYS_PER_YEAR = 365;
    const int DAYS_PER_LEAP_YEAR = 366;
    const int DAYS_PER_LEAP_CYCLE = 1461; // 4 years
    const int JULIAN_TO_GREGORIAN_THRESHOLD = 2299161; // Oct 15, 1582

    // Adjust for Julian-to-Gregorian transition
    if (d >= JULIAN_TO_GREGORIAN_THRESHOLD) {
        d += 10; // Account for dropped days
    }

    int a = d + 32044;
    int b = (4 * a + 3) / DAYS_PER_LEAP_CYCLE;
    int c = (4 * a + 3) % DAYS_PER_LEAP_CYCLE;

    int y = (b / 1460) + 1970; 
    d = (c / 4) - 365;

    if (d < 0) {
        y--;
        d += DAYS_PER_YEAR + (y % 4 == 0);
    }

    std::cout << "Gregorian Year: " << y << std::endl;
    std::cout << "Day of Year: " << d + 1 << std::endl; // Add 1 as days are typically 1-indexed

    return 0;
}

Not only that, but you can ask it to rewrite it in iambic pentameter and with pirate-speak identifiers.

Really begs the question of what the long-term outlook is for the non-top-quartile people. Maybe re-prompting LLMs is the new ball of mud?

> Maybe re-prompting LLMs is the new ball of mud?

Recently, my company acquired some engineers, and while I cannot evaluate their "percentile", I have been low-key flabbergasted at some of their suggestions for using LLMs in place of scripts. (For example, to modify or move nested data in a non-public JSON representation of a program.)

> As a rule of thumb, leap days come around every four years. But there are exceptions to this rule. For example, at the turn of every century we miss a leap year. Even though the year is divisible by four, we don't add a leap day in the years that end in 00. But there's an exception to this rule too. If the year is a multiple of 400 then we do add in an extra leap day again. At the turn of the millennium, despite being divisible by 100, the year 2000 did, in fact, have a 29 February because it was also divisible by 400.

https://www.bbc.com/future/article/20240228-leap-year-the-im...

That reminds me of this program, which calculates the month and day given the year and the day of the year [0]. There's a version that abuses goto, and one that stores state in variables and only uses structured statements. You probably won't see people similarly abuse goto today, though. I think it was translated from Fortran.

[0]: https://craftofcoding.wordpress.com/2020/02/12/the-world-of-...

Too simple. https://www.bbc.com/future/article/20240228-leap-year-the-im...

Not for the task at hand. The initial d was divided down from an unsigned 32-bit timestamp in seconds from the Unix epoch (hence the 1970.) It accounts for 2000 (leap) and 2100 (not leap), after which the u32 overflows.

I am not talking about ease of programming, and c being imperative, and the explicit break are edge cases.

But for case value

value can only be char or int, and must be unique

Really it is just sugar for if-else-if, but if you enforce explicit break in your code it is as close to a provable total function as you get in C.

Total functions are the ideal in any code.

As determining whether or not a function F is total is undecidable, switch is more reliable than if-else-if ladders.

Was it then case that there was special logic attached to special leap years like all of the exceptions we hear about in other software products like Excel etc? Or was it just "return true" and "return false" ?

At least you can be sure that it worked, whereas if he'd actually written the succinct version there was a risk of a subtle bug.

Why would such a system assure someone that it worked? Conversely, why would only the succinct version lend itself to a subtle bug?

Mine is as succinct as they come, but a little bit buggy

  int is_leap_year(int year) {
    return 0;
  }

I don't think that qualifies as a "subtle" bug. :)

Suppose that someone tried to naively derive the leap year by dividing the current year by 4. There are some years that are divisible by 4 but that are nonetheless not leap years. If you fail to account for this in your succinct implementation then you might introduce a leap day that doesn't exist at some point. I know the table implementation seems really dumb, and it's not that hard to look up a correct algorithm for this, but it's a good example. If you build a table from some correct source then you can simply use that table instead of running a calculation. There are fewer than 30 leap years in the next 100 years, by which time something cataclysmic will happen that will completely erase the need for your code anyway. At least, you will be dead before you need to revise the code again. So it's reasonable to build a function with a table given the time-limited nature of the code and the desire to be visibly and explicitly correct. Sure you could add a unit test with the table and test the calculation, but that actually adds more code to maintain than just jamming the table directly into the function.

I don't think they were stating that a simple table based solution was used, if it were, it would probably meet the bar for "succinct" because you could just write that as something like `if X in Y`. Rather, I got the impression that a lookup based approach was intended, but it was enterprise-ified (like https://github.com/Hello-World-EE/Java-Hello-World-Enterpris...).

Oh, gross.

Doing modulo 4, 100, and 400 isn't really that complicated to implement. I know there are even simpler ways to calculate this but come on, this is stupidly easy stuff. Then again we live in a dimension where there are Node packages for IsEven() and IsOdd() with tons of downloads.

Sure, in Node.JS, and with a specific calibur of programmer in a specific business context this is easy. Technical problems are easy to solve, but people and organizational problems are where the real challenges of engineering come. Sometimes you have to grit your teeth and implement something for the company you work for rather than for your own personal sense of correct.

You can only be sure if you verified every year and didn't miss a typo. Having a test that straightforwardly checks every year individually would be more useful, since then you know if you changed anything, and can double-check just the changes.

Although, I suppose it doesn't matter all that much which is which, as long as they're not both written the same way.

How do you ensure that you don't put a typo in the unit test?

Hopefully, it wouldn't match the algorithm.

It's implementing it two different ways and comparing the results. It's not a guarantee, since you could have implemented the wrong spec, but there's only so much you can do with internal consistency checks.

But it would be better to compare to a known good function. It might be easier to add a dependency to do that in a test than in production code.

If you used the list of years in the test and the mathematical expression in the impl (or vice versa), then you know you didn't make the same mistake in both.

Unless there's something about leap years that you don't understand and miss in both tests and implementation. Not an impossible situation.

Reply on news.ycombinator.com