Ask Slashdot: What Tools To Clean Up a Large C/C++ Project?

An anonymous reader writes I find myself in the uncomfortable position of having to clean up a relatively large C/C++ project. We are talking ~200 files, 11MB of source code, 220K lines of code. A superficial glance shows that there are a lot of functions that seem to be doing the same things, a lot of 'unused' stuff, and a lot of inconsistency between what is declared in .h files and what is implemented in the corresponding .cpp files. Are there any tools that will help me catalog this mess and make it easier for me to locate/erase unused things, clean up .h files, and find functions with similar names?

Comments

rm

[Anonymous Coward]

Who about "rm"?

Re:rm

[fahrbot-bot]

Who about "rm"?

Ah yes. Every *nix programmer has, hopefully only once, experienced the joy of the following:

% rm * .o
.o: No such file or directory

Static analysis tools...

[underqualified]

If you're company is willing to pay for it, you can get something like Coverity. On the free(as in beer) side there is CppCheck and clang.

You call that large?

[Anonymous Coward]

Seriously, that's mid-sized at best.

Re:You call that large?

[Oxdeadface]

What compels comments like this? The first AC posts absolutely nothing of value, just wants to let everyone know that they disagree with a minor point that's completely irrelevant to the OP's question. Thanks for the insight, champ. The followup, probably the same person, goes on to ramble like an old fart telling a useless anecdote about his kid that's barely even related to the topic at hand. At what point did either of these seem like a good idea? Neither of these comments address the question being asked or even attempt to be useful at all. No one cares what you consider a large program and absolutely no one gives a shit about you or your fucking crotch fruit. These comments are just some sad cunt's way of claiming, "I'm more experienced and better than you." Fuck right off.

clang static code analysis

[Anonymous Coward]

scan-build and scan-view from clang++ will show you what is being used and what isn't as far as static code analysis goes.

Document first

[gbjbaanb]

So, figure out the layers or logical components between each module and then you will be able to chew smaller chunks.

Then, doxygen the whole lot, making sure to use dot to create the graphs for callers and callees. This will let you see the interaction points so you can see what impact a change in one method will have (ie which callers you have to check).

Some people will say "write unit tests" but frankly, it never works with a legacy code base, to effectively unit test you have to write your code differently to how you'd normally do it. You don't have that luxury here. So a good integration test suite should be developed to test the functionality of the whole thing, then you can repeat it to make sure your changes still work. Its not as instant as unit testing (but more effective) so you'll have to invest in a build system that regularly builds and runs the (automated) integration test and tells you the results - and commit changes reasonably regularly so you can isolate changes that end up breaking the system.

The rest of the task is simply hard work running through how it works and understanding it. There's no short-cuts to working hard, sorry.

Re:Document first

[bmajik]

This.

One of my first professional programming projects was to take a look at the custom C++ billing software our company had purchased from a contract programmer.

I had a long unix and programming background, and was back for a summer job after doing 1 semester of C++ in college.

My boss told me, since I was the only one who had C++ experience, to start documenting the system.

At the time, we were using IRIX, and so I was using the SGI compiler and tools suite, which were, I believe, licensed from EDG. The point is that there was a very nice call graph visualizer. This was helpful for understanding things at a superficial level.

However, what was even better was just running the program a bunch of times on test data and seeing what it did while under the debugger.

While my summer began with the task of documenting the system, as I learned things I'd report them to my boss.

By the end of the summer, I had re-written some fundamental parts of the system; I'd moved some of the processing outside, and I pre-processed and pre-sorted the data.

The overall execution time went from many hours to about 45 minutes to calculate monthly bills. THe key innovation was replacing the inner loop of the charge tabulation -- which was 2 or 3 levels of nested linked list traversal.

Instead, I used the standard unix sort tools to pre-sort the data files before being loaded into the system, and I changed the system to use a data structure that supported a binary search.

The majority of the code got left alone. By understanding the code under a debugger, and realizing that how it worked on production data was much different than how it performed on the test data it was originally delivered with, I was able to make a critical set of changes that had a huge impact.

In general, I spend as much time as I can not writing code, but instead, understanding how the existing system works. For a current project, I've spent the last two weeks playing with somebody else's code, and now I've expanded it so that it can also operate on my data sets, and I've probably changed fewer than 100 lines across about 5 different projects.

Eclipse, Xcode or any IDE

[guruevi]

Any decent IDE has the capability of pointing at least towards unused blocks of code and will generate a tree of function calls. I've worked with Eclipse and Xcode both of which have these capabilities. Even GCC (or another C compiler) can warn you about chunks of unused code or missing/bad header files. You can also rename functions across the entire codebase if necessary.

If your code has warnings or errors, continue fixing until the warnings are gone. As far as functions that do similar things but are named differently, that is a bit harder because 'looks like they are doing the same thing' doesn't always mean they ARE doing the same thing (if they have the exact same code, you could perhaps solve with statistical analysis or simply a text finder).

Make sure that if you replace a function that it has the same behavior in all cases. Even mediocre developers have learned that reuse existing code is a "good thing" and often different functions that do "the same thing" have edge cases (often undocumented) where it does behave differently (especially in C/C++ eg. difference in signedness, memory mapping method, characters etc)

If you don't know what it does, don't touch it.

[BlueKitties]

Seriously, you never know when some previous programmed made a "duplicate" function to do something bizarre, like force a particular initialization order of static-class-member variables between translation units. Sometimes deleting pointless code can do... terrible things. Just be careful, test your changes, etc.

Unit tests

[Midnight Thunder]

While I dislike writing unit tests, I have to admit they are useful in protecting your butt when something breaks, since the test should catch it first. Of course you need to decide whether in a particular scenario they add value or just make you manager happy.

In a case like yours, you can make code modifications and hope nothing breaks or build unit tests and ensure that you don't break any of them when refactoring. Initially rather than just ripping out the seemingly duplicate methods, rip out/tweak their implementation and have them point to what they seems like a the right method to provide the common functionality. If your unit tests show breakage, then you know that you missed something.

If you do things wholesale, then you are likely to break something in an unmanageable way. Oh and make sure things are version controlled ;)

Re:Unit tests

[gstoddart]

I've maintained several legacy code bases over the years.

And I will flat out tell you that unit tests have VERY limited utility in terms of understanding a mess of code you inherited. At least, in the beginning.

Sure, you can start with a couple of basic premises, and you can convince yourself those basic premises still work.

But the initial grokking of your code, understanding all places where a function may be used, understanding all of the tricky bits and gotchas, trying to understand why there are 9 functions which look like they do the same thing? That takes some time and effort, and quite possibly some tools.

Unit tests are great for starting to build up a few things, and move towards better stuff ... but in a system which has several hundred (or several thousand) functions and interactions, resulting in really large numbers of code paths ... having a few unit tests describing the stuff you understand doesn't mean all of the stuff you don't understand wasn't broken, simply because you don't know what you don't know.

So it is important to understand your new unit tests on legacy code are, at best, a VERY incomplete view of your code. That will improve over time, but you could potentially need to write a few thousand of them to be sure you're not breaking anything in the big picture.

If you do things wholesale, then you are likely to break something in an unmanageable way. Oh and make sure things are version controlled ;)

Oh, yes .... This .. for the love of god, this.

You should learn how to tag branches and the like in your version control so you can identify a baseline of "before I ever touched anything" and then be able to cleanly build everything which predates you, as well as building your "after refactoring this part".

Branching/tags/whatever your version control calls it -- that doesn't take up much space, so use them often, and consistently. Let the tool do the heavy lifting of keeping track of what you've changed.

You do NOT want to find yourself unable to build it as it existed, or identify all of the diffs between what you started with and what you have.

Looks like a reverse engineering project

[prefec2]

Modularize the software. There are a lot of tools which can help you to analyze static dependencies in the code which can help you to identify components. You could also use a run-time analysis tool for example Kieker which is initially for Java, but there is an extension for C/C++.

Few ideas

[postmortem]

1. Modern IDE with good gcc parser: Eclipse, Netbeans, 3rd party paid ones. Not Visual Studio. You want it to build call hierarchy tree for you, so that you can find methods that are unused. It will require some manual steps
1a. if you have $, Understand for C/C++ is proprietary tool that will map a hierarchy of your code.
2. perform structural coverage analysis of code in live action, will help map the dead code. gcov is free if you can use it.

Re:Does Lint Exist anymore

[OrangeTide]

Compiler warnings have mostly caught up with the capabilities of Lint. There are some things Lint still does, but there are lots of things it warns about that have, as far as I know, never been the cause of a real bug. Getting a project to be 100% warning free with gcc -Wall is possible, and usually possible with -Wextra (maybe not so much with g++). The warnings usually are valuable, and I've personally seen bugs that could have been caught with gcc's warnings. Other compilers have other warnings and personalities, but I think it's worthwhile to investigate using warnings to check out a project with any compiler.

Answer: read slashdot for long enough

[plcurechax]

See: Working Effectively with Legacy Code book review (2008) for a book of that title by Michael Feathers (PDF article) on that very topic.

There is even a summary of key points at Programmers @ StackExchange. Hundreds if not thousands of programmer's blogs address this very topic.

You're welcome. Now get back to work.

DXR, the code indexer

[Grincho]

Wow, what an easy pitch. :-) At Mozilla, we've put together a tool called DXR ( https://github.com/mozilla/dxr... ). It indexes your code and lets you do text and regex searches. But if you can get your project to build under clang, you can really have some fun, with queries that find...

* Calls of a function (great for dead code removal)
* Uses a type
* Overrides of a method
* Uses and definitions of macros
* etc., etc., etc. There are something like 24 different structural queries you can do.

Because all of this is informed by the internal data structures of the clang compiler, it's nigh on 100% accurate (aside from more dynamic behaviors like sticking function pointers in a table and passing them around). You can also explore a hyperlinked version of the source, bouncing from #include to #include and drilling into methods.

Here's how to set it up: https://dxr.readthedocs.org/en...
Here's our production instance you can play with: https://dxr.mozilla.org/mozill...

If you run into trouble, pop into #static on irc.mozilla.org, and we'll be happy to help you.

I've done this far too many times

[WinstonWolfIT]

First off, 220k lines of source isn't that big.

You're not going to solve this with a big bang so get that idea out of your head. You're going to solve it gradually, and for a code base of that size it's going to take maybe a year of relatively slow improvement. Everyone on the team has to be on board, and every code review must include "What has been improved?" and "Did anything get worse? If so, that's not okay."

1) Pick your battles. The code you're not changing is code that doesn't need to be looked at. Address your pain points as they come up.
2) When you find a pain point while making a change, MAKE IT TESTABLE. Since you're in here making a usually simple fix, a single nominal test verifying that fix is fine. Testing anything else is a waste of time. Testable code will improve over time.
3) If you can't make code testable because of an intractable dependency graph, welcome to the hell of "Design Dead". The only way out of this scenario is a rewrite of that area.
4) Find your comfort level with regard to time boxing refactoring work. On my engagements, they just happen automatically, without explanation outside the team, nor apology to anyone. When estimating a piece of work, pad it with some extra time for cleanup. Only actually create work items for design dead areas. Your definition of done must include testable, tested and improved code.
5) Duplicate code in itself isn't evil, and inconsistencies are simply inevitable. If you find duplicate code, pick one and deprecate the rest. However, code that is tightly coupled to the deprecated code will need to be refactored and if the coupling traverses an extended dependency graph, you'll simply have to live with the duplication and just stop adding to it.

Bware of 'cleanups'

[plopez]

Anecdote from the mists of time:

There was this C program which had been around a while which had undergone some evolution and maintenance. The decision was made to 'clean it up' There was a data structure, an array I think, which was unused in a subroutine, lets call it subroutine A. So it was removed. The next test runs of the application and suddenly the program started core dumping. After some agonizing debugging it was discovered to come from another subroutine, lets call it subroutine B.

There had been an array in subroutine B which a loop had run over the end of. But subroutine A had loaded just prior to B and allocated memory for the unused data structure. This had provided enough space to handle the array out of bounds error in subroutine B but when removed subroutine B began overwriting subroutine A causing the crashes.

It was good that the crashes were easily reproducible or could have been one of those intermittent things that drive people insane. An automated tool may not catch things like that since it may not show up until run time. It is C/C++ we are talking about now isn't it?

Re:But are you lacking experience and the brain fo

[I'm New Around Here]

Hey, MC Hammer built my house for me.

Unfortunately, I'm not allowed to touch it.

Comment removed

[account_deleted]

Comment removed based on user account deletion

Could it be a threading issue like a a deadlock?

[Paul Fernhout]

Debugging code that prints or logs may act to synchronize access to some data structure. Sometimes that can prevent a deadlock or illegal pointer access as a side effect:
http://stackoverflow.com/quest...
http://en.wikipedia.org/wiki/D...

So yes, complex programs can act in strange ways from seemingly minor changes.

I spent a couple years helping maintain a large complex multi-threaded app (which included message passing between the apps, for another layer of fun) which supported 24X7 operations where a minute's downtime could cost millions of dollars in some situations, and it was not easy. The code base was easily 10X to 100X of what the poster of the story is tasked with maintaining. Versions of the code had been in production for over fifteen years. Much of the code had been ported from C++ & Tcl to Java (although C++/Tcl systems remained), but the threading model was somewhat different between the two, and the port had not taken account of all the differences. It would have been nice to be able to rewrite some key parts of the system to make them more maintainable, but there was never enough time for that in a big way -- and realistically, bigger rewrites likely introduce new issues. Still, eventually we got most of the worst deadlocks and memory leaks and similar such things fixed and the system got to the point where people stopped even remembering off-hand the last time a core part of the system needed to be rebooted (previously a fairly frequent event). But each deadlock could involve days, weeks, or even months of study and discussion, adding log statements, writing tests, lab tests, analyzing quite a few multi-gigabyte log files (and writing tools to help with that including visualizing internal message flow), and so on. And, same as you mention, hardware and OS issues could interact with it all, making some things hard to duplicate under virtual machines for developers. One thing is that to the end user, a system that is more stable may not look that different than one that is less so -- there are no new features, so it is not obvious what is being paid for.

Although obviously if the program you support core dumps from a bad address or stack overflow, rather than just freezes up, it is probably something else. Still, even then, a bad pointer address can sometimes come from one thread freeing a data structure when another thread is still using it. The original C++ in the above mentioned project generally was highly reliable, but it still had some odd issues too. In one rare case, memory was freed in an unexpected way under certain conditions by other code running in the same thread but in code nested way deep with essentially recursive calls processing complex messages. I finally also traced part of that too what looked like maybe a bug in a supporting third-party library (a RogueWave data structure). Because that C++ code had been in production for years, and we were loathe to change it at the risk of introducing new issues, we mostly "fixed" that issue by making changes elsewhere in the system to prevent that component from getting the pattern of data that it had trouble handling. But we would not have known exactly what to change elsewhere without a lot of analysis.

Sadly, just as we got it mostly working well, the new shiny thing of a mostly COTS system that did something similar came along to replace much of it (at a much bigger expense than maintaining the old, but granted with some nice new features).

As I saw someone else comment recently about a "stable" OS, the end user generally cares more about how much work a system lets them get done, not how "stable" it is. A reboot can be acceptable, depending on the situation and the alternatives, even if not desirable. Erlang code is probably the master at that approach of rebooting code when it fails. :-) Here