Increased analysis time since SVF 2.5

[mortkever]

Hello,

I recently bit the bullet and migrated from version 2.5 to 3.2. However it seems to me the analysis takes a lot longer to run.
The difference is most notable in 483.Xalancbmk from the SPEC2006 bench marking suite. A simple Andersen wave diff takes about 20 seconds in version 2.5 and close to 12 hours in 3.2. This with 96GBs of RAM and a Ryzen 9 9950X3D, so I doubt my setup is suddenly the bottleneck.
I assumed it was due to some mechanism that increases precision which is currently not a priority for my project.
However I also set the "MaxFieldLimit" to zero to disable field sensitivity and "ContextInsensitive" to true, which unfortunately made little difference.

The increase in duration seemed so disproportionate, it warranted an issue.

Are there other options that were introduced that greatly increase the analysis time? Are there other ways to increase the performance or reduce the time needed?

Thanks in advance!

[yuleisui]

I believe this is because SVF has evolved over the years with improved soundness for C++ programs such as 483.xalancbmk, which makes extensive use of STL library functions and virtual calls. As a result, Andersen’s analysis can become overly conservative in this context.

You could try reducing the limit for resolved indirect calls, for example by setting -ind-call-limit=100.

SVF/svf/lib/Util/Options.cpp

Line 291 in 2060a61

"ind-call-limit",

We also have a Simplified-STL project to reduce the analysis overhead on existing STL lib functions, which are not necessarily to be analyzed if you focus on application code. You could have a try, but this is a side project that has not been actively maintained.

https://github.com/SVF-tools/Simplified-STL