Optimization considerations#

The performance of angr as an analysis tool or emulator is greatly handicapped by the fact that lots of it is written in Python. Regardless, there are a lot of optimizations and tweaks you can use to make angr faster and lighter.

General speed tips#

  • Use pypy. Pypy is an alternate Python interpreter that performs optimized jitting of Python code. In our tests, it’s a 10x speedup out of the box.

  • Only use the SimEngine mixins that you need. SimEngine uses a mixin model which allows you to add and remove features by constructing new classes. The default engine mixes in every possible features, and the consequence of that is that it is slower than it needs to be. Look at the definition for UberEngine (the default SimEngine), copy its declaration, and remove all the base classes which provide features you don’t need.

  • Don’t load shared libraries unless you need them. The default setting in angr is to try at all costs to find shared libraries that are compatible with the binary you’ve loaded, including loading them straight out of your OS libraries. This can complicate things in a lot of scenarios. If you’re performing an analysis that’s anything more abstract than bare-bones symbolic execution, ESPECIALLY control-flow graph construction, you might want to make the tradeoff of sacrificing accuracy for tractability. angr does a reasonable job of making sane things happen when library calls to functions that don’t exist try to happen.

  • Use hooking and SimProcedures. If you’re enabling shared libraries, then you definitely want to have SimProcedures written for any complicated library function you’re jumping into. If there’s no autonomy requirement for this project, you can often isolate individual problem spots where analysis hangs up and summarize them with a hook.

  • Use SimInspect. SimInspect is the most underused and one of the most powerful features of angr. You can hook and modify almost any behavior of angr, including memory index resolution (which is often the slowest part of any angr analysis).

  • Write a concretization strategy. A more powerful solution to the problem of memory index resolution is a concretization strategy.

  • Use the Replacement Solver. You can enable it with the angr.options.REPLACEMENT_SOLVER state option. The replacement solver allows you to specify AST replacements that are applied at solve-time. If you add replacements so that all symbolic data is replaced with concrete data when it comes time to do the solve, the runtime is greatly reduced. The API for adding a replacement is state.se._solver.add_replacement(old, new). The replacement solver is a bit finicky, so there are some gotchas, but it’ll definitely help.

If you’re performing lots of concrete or partially-concrete execution#

  • Use the unicorn engine. If you have unicorn engine installed, angr can be built to take advantage of it for concrete emulation. To enable it, add the options in the set angr.options.unicorn to your state. Keep in mind that while most items under angr.options are individual options, angr.options.unicorn is a bundle of options, and is thus a set. NOTE: At time of writing the official version of unicorn engine will not work with angr - we have a lot of patches to it to make it work well with angr. They’re all pending pull requests at this time, so sit tight. If you’re really impatient, ping us about uploading our fork!

  • Enable fast memory and fast registers. The state options angr.options.FAST_MEMORY and angr.options.FAST_REGISTERS will do this. These will switch the memory/registers over to a less intensive memory model that sacrifices accuracy for speed. TODO: document the specific sacrifices. Should be safe for mostly concrete access though. NOTE: not compatible with concretization strategies.

  • Concretize your input ahead of time. This is the approach taken by driller. When creating a state with entry_state or the like, you can create a SimFile filled with symbolic data, pass it to the initialization function as an argument entry_state(..., stdin=my_simfile), and then constrain the symbolic data in the SimFile to what you want the input to be. If you don’t require any tracking of the data coming from stdin, you can forego the symbolic part and just fill it with concrete data. If there are other sources of input besides standard input, do the same for those.

  • Use the afterburner. While using unicorn, if you add the UNICORN_THRESHOLD_CONCRETIZATION state option, angr will accept thresholds after which it causes symbolic values to be concretized so that execution can spend more time in Unicorn. Specifically, the following thresholds exist:

    • state.unicorn.concretization_threshold_memory - this is the number of times a symbolic variable, stored in memory, is allowed to kick execution out of Unicorn before it is forcefully concretized and forced into Unicorn anyways.

    • state.unicorn.concretization_threshold_registers - this is the number of times a symbolic variable, stored in a register, is allowed to kick execution out of Unicorn before it is forcefully concretized and forced into Unicorn anyways.

    • state.unicorn.concretization_threshold_instruction - this is the number of times that any given instruction can force execution out of Unicorn (by running into symbolic data) before any symbolic data encountered at that instruction is concretized to force execution into Unicorn.

    You can get further control of what is and isn’t concretized with the following sets:

    • state.unicorn.always_concretize - a set of variable names that will always be concretized to force execution into unicorn (in fact, the memory and register thresholds just end up causing variables to be added to this list).

    • state.unicorn.never_concretize - a set of variable names that will never be concretized and forced into Unicorn under any condition.

    • state.unicorn.concretize_at - a set of instruction addresses at which data should be concretized and forced into Unicorn. The instruction threshold causes addresses to be added to this set.

    Once something is concretized with the afterburner, you will lose track of that variable. The state will still be consistent, but you’ll lose dependencies, as the stuff that comes out of Unicorn is just concrete bits with no memory of what variables they came from. Still, this might be worth it for the speed in some cases, if you know what you want to (or do not want to) concretize.

Memory optimization#

The golden rule for memory optimization is to make sure you’re not keeping any references to data you don’t care about anymore, especially related to states which have been left behind. If you find yourself running out of memory during analysis, the first thing you want to do is make sure you haven’t caused a state explosion, meaning that the analysis is accumulating program states too quickly. If the state count is in control, then you can start looking for reference leaks. A good tool to do this with is https://github.com/rhelmot/dumpsterdiver, which gives you an interactive prompt for exploring the reference graph of a Python process.

One specific consideration that should be made when analyzing programs with very long paths is that the state history is designed to accumulate data infinitely. This is less of a problem than it could be because the data is stored in a smart tree structure and never copied, but it will accumulate infinitely. To downsize a state’s history and free all data related to old steps, call state.history.trim().

One particularly problematic member of the history dataset is the basic block trace and the stack pointer trace. When using unicorn engine, these lists of ints can become huge very very quickly. To disable unicorn’s capture of ip and sp data, remove the state options UNICORN_TRACK_BBL_ADDRS and UNICORN_TRACK_STACK_POINTERS.