Things I learned the hard way

[Exil]

On Thursday, 27 June 2019 at 09:10:36 UTC, Gregor Mückl wrote:
> On Wednesday, 19 June 2019 at 02:41:31 UTC, Exil wrote:
>> On Wednesday, 19 June 2019 at 01:15:33 UTC, Walter Bright 
>> wrote:
>>> https://blog.juliobiason.net/thoughts/things-i-learnt-the-hard-way/
>>>
>>> Debuggers are over-rated
>>> I heard a lot of people complaining that code editors that 
>>> don't come with debugging are terrible, exactly because they 
>>> don't come with debugging.
>>> 
>>> But when your code is in production, you can't run your 
>>> favorite debugger. Heck, you can't even run your favourite 
>>> IDE. But logging... Logging runs everywhere. You may not have 
>>> the information you want at the time of the crash (different 
>>> logging levels, for example) but you can enable logging to 
>>> figure out something later.
>>> 
>>> (Not saying debuggers are bad, they just not as helpful as 
>>> most people would think.)
>>
>> This I don't agree with. Maybe specifically for his job it 
>> might not have been. If you are more of a tech support and you 
>> need to remote into a customer's computer and use whatever 
>> they have available on their computer, sure. But I doubt 
>> that's the case for a large majority of programmers. Also you 
>> can debug a crash that happened on a customer's computer for a 
>> native production application with a stack trace. You might be 
>> missing some variables because it is a production build that 
>> were optimized away but you still get a lot of relevant 
>> information.
>
> I think you're looking at this paragraoh from the wrong angle. 
> For me it is more about the skill of debugging than about the 
> availability of a debugging tool. I would rewrite it as "learn 
> how to debug code in your head". It makes more sense that way: 
> be sure that you can reason about your code's behavior reliably 
> without tool support. This includes keeping things as simple as 
> possible and knowing how stuff works, including stuff you 
> didn't write.
>
> So, for example the nifty container class you're relying on may 
> make things simpler to code, but maybe you're hitting a usage 
> pattern that the code on the other side of the container 
> interface doesn't like. In C++, this means e.g. knowing exactly 
> when a std::vector might reallocate its backing memory. And 
> that might be the reason that the one reference you kept from 
> before the push_back call might no longer be valid. Figuring 
> out things like that with a debugger is usually hard because of 
> the way most STL implementations are written. Their internals 
> might be suppresses by the debugger, show up as a convoluted 
> chain of method calls, or be optimized so much that you can't 
> make heads or tails of it.
>
> So, bottom line: even if you have a debugger, don't count on 
> it. If you get that right, this make you a better programmer 
> and improves your code.
>
> Enough preaching to the choir.

I don't think he said it in that manner, he was specifically 
talking about the debugger. When a crash happens it gives you 
information about the crash. So it just provides a means of 
getting extra information to you to help you diagnose the problem.

If you only used logging as a means to diagnose the problem, if 
you do have that kind of error. At least with the debug 
information you can see the problem is in push_back. You can see 
what the data is. Even if it is optimized, you will still at the 
very least get the line number and the call stack of where the 
crash happened.

Then there's bad code gen, it doesn't matter how good you are at 
debugging in your head, if the problem isn't even with your code. 
I think I've seen bad code gen in C++ once or twice, if even 
that. With DMD on the other hand, there is routinely bad code 
gen. Just a little while ago, boolean has code gen that can cause 
it to be interpreted either way if it isn't initialized. The 
`bool b=void; if(!b){}` I imagine that code is literally 
converted into a negation instead of just using JNE instead of JE 
for the comparison. That's not mentioning all the problems with 
C++ interpolation I've experienced.

You should use the tools available to you, the debug info from a 
crash just serves to give you additional knowledge. Pretending 
it's not there doesn't make you a better programmer, it just 
gives you less information to work with. Depending on what the 
problem is, like someone else mentioned with null dereferencing, 
it can be something you solve in 5 mins with debug info, or 
almost impossible depending on how big your code base is and how 
detailed your logging is with the performance requirements.