doxygen runtime went from 10minutes to >14hours (lots of `using namespace`)

[martinus]

Describe the bug
Newest doxygen version is impracticably slow when there are lots of using namespace in the code. I stopped generating after 14 hours. Older versions take 10 minutes for the same codebase.

Expected behavior
Don't take forever

To Reproduce
Unfortunately I cannot share our code. It might be reproducible with several files and lots of namespaces, I can try to do that if you think that helps

Version
I checked out the latest version, and did a git bisect to find out when the slowdown was introduced. It first happened in this commit: 8a93f4e. Before this commit a stripped down codebase takes 0.5 seconds to generate, after that commit I stopped it after several minutes. I did all my tests in Linux, with clang++

Stack trace
I used hotspot to figure out where doxygen was hanging (for this I built master), and got this flamegraph:

Additional context
Looking at the changes in 8a93f4e I suspect the issue is now too often a fresh AccessStack or VisitedNamespaces or StringUnorderedSet is used when previously a non-empty container was used

[albert-github]

You pointed to a probable cause of the problem (thanks), though you also wrote:

Newest doxygen version is impracticably slow when there are lots of using namespace in the code. I stopped generating after 14 hours. Older versions take 10 minutes for the same codebase.

• Which is the "Newest doxygen version" version you tried (output of doxygen -v)
• Which is the most recent "Older versions" you tried?

especially the "Newest doxygen version" is of important.

The source code you used:

• which language is it written in?
• is it public domain or propriety?

[martinus]

Which is the "Newest doxygen version" version you tried (output of doxygen -v)

I've built doxygen myself, this is the output of doxygen -v of the newest version that I've tried:

1.9.4 (de4a15d)

Which is the most recent "Older versions" you tried?

The most recent old version (last good version) I tried:

1.9.0 (02cdddb)

Then, first bad version is this:

1.9.0 (8a93f4e)

The source code you used:

It's C++, and proprietary.

[doxygen]

@martinus It would be very useful to have a dummy example, modelled after the proprietary code if needed, that allows me to reproduce the problem. It is enough if it shows a significant (say >2x) performance degradation.

[martinus]

I'm trying to create a stripped down example, but it seems to be a bit difficult to narrow it down. I see what I can do

[ferdymercury]

Not sure if related #5875

[martinus]

Sorry, it took me a long time to create a stripped down example that demonstrates slowness. In the end I have now this repository, which basically 2 files that have lots of using namespace: https://github.com/martinus/doxygen_slow_namespace_example

Generating doxygen for the ~80 lines of C++ code takes 1m 34s on my machine. It seems thow that the old version 1.9.0 is also slow on these files, the binary from 02cdddb takes 47 seconds.

[albert-github]

Thanks for the example, but why a repository (which is potential non persistent) and nut just a simple tar or zip file like: example.tar.gz

[albert-github]

It looks like it is related to the fixed problem #6382 but apparently we have her another case.

I ran with doxygen -d time (and set QUIET=NO) and see as potential place:

0.073 sec: Generating file sources...
0.074 sec: Generating code for file src/1.h...
531.519 sec: Generating file documentation...

[doxygen]

@martinus Thanks for the example, it should definitely not slow down things so much, so I'll look into it. That said, I see it uses "using namespace" in a header. Did you know this is considered a bad practice?

[martinus]

@doxygen Thank you for looking at this! You are right, using namespace in header is definitely bad practice, that code example I posted is not an indication of my programming style 😆 It's just what I got from trying to reduce a large codebase down to a simple example to demonstrate the issue