BinNavi's basic block handling

A while back I talked about the problem of highly optimized code and the resulting problems when we want to store it in a database while allowing for all possible constructs.
In this post I'll show how the recently released BinNavi 1.2 handles some cases where code is shared among functions and basic blocks exhibit non-typical characteristics.

Note: I could not get the Microsoft or Intel compilers to produce compiled code with functions sharing basic blocks as an example for this post. Using PGO (Profiling Guided Optimization) and other optimizations proved fun but the furthest I could go was producing multi-chunked functions. I might need to play more with it in order to get chunks to be shared... Anyway, I just picked a Microsoft DLL for the example. Specifically wkssvc.dll.

In this DLL there are two functions sharing some blocks on their exit paths. The shared code is shown in green in this graphic exported from BinNavi.





Interestingly enough, those blocks share exactly the same code but one function has six of them while the other has seven. I previously commented on the issue here and happends because of a reference in one of the functions that targets the shared code and causes a block to split.

Here one can see the shared code grouped and highlighted. Click on the image for a larger view.



Now a zoomed version of the block on the left:



And the one on the right:



Within the database, BinNavi handles this by allowing instructions to belong to multiple basic blocks, as well as basic blocks to belong to multiple functions. And it just works...

Multi-chunked functions and IDA

In my post a few days back I did mention the problem with shared basic blocks across functions and made the assumption that IDA could not handle them. It is the case, as clarified by Ilfak, that IDA can actually handle them, yet they are not currently assigned to multiple functions. IDA's plugin API even allows to iterate through the chunk's parent functions.

Simply blocks, basically...

A few days ago I bumped into something I was not really counting on seeing. Compiler optimizations have surely gone a long way.

Some background first...

After having seeing functions split over non-contiguous basic blocks for some time now, it was quite natural to think that some of those basic blocks could be shared among functions ( obviously, only the ones leading to the function exit points as once the shared code is reached, there's no way of getting flow back to basic blocks not shared by those functions).

Then we have that functions can be split with their blocks in different parts of a binary and some of those blocks shared. The reason for the splitting comes from doing profiling in normal use-cases of the applications and trying to group frequently accessed code into as few pages of the executable as possible, so that a minimum set of those need to be mapped at one time in memory. Only when infrequently visited code is reached some new pages new to be mapped. The following figures illustrates the concept.

UPDATE: Just got told that the reason for the splitting is more likely to be there to take advantage of the internal CPU instruction cache than of memory paging. Keeping the frequently traversed code together will result in less instructions being fetched from RAM (slower) for that code area. Also will allow to fit more code in the code-cache by moving away the less used blocks.

Here we can see the blocks being laid out continuously in memory. As can be normally seen in non-optimized code.



This would be how the same function would be laid out if profiling information is incorporated, so that frequently traversed paths are together within the code (in the same memory page if possible, in order to reduce memory footprint and paging).




Once one has the splitting, the idea of sharing comes naturally.

This results in that, from the disassembler point of view, one has to allow for those chunks and also for those chunks to be assigned to an arbitrary number of "owning" or parent functions.



What is more interesting, and the subject of this post, is the fact that instructions can also belong to different basic blocks. At least, under one view. This arises from cases where extensive optimizations are used.

A couple of days ago I was looking into an optimized binary (the craziest I have seen in a while) and how it was mapping into the SQL representation we are using at Sabre, there were some problems when exporting the information from IDA. (IDA can't really handle too well (yet) heavily-"chunked" code, so I have to account for that and build intelligence that analyzes the code for cases like the one I'm discussing here)

The problem was with two functions sharing a number of basic blocks, the funny side was that, depending which function one analyzes the flow among the shared blocks will look different. And the cause is fairly obvious too once one realizes why the problem appears.

A conditional branch from the non-shared code in one of the functions targeting the shared code will cause a split in the flow. A split which is not present from the other function's point of view. The following figure shows the result of a branch into shared code from only one of the sharing functions.



There are two solutions for this problem. One would be to represent the same basic blocks all over the binary, which would introduce a non-natural spit in a function, the other way would be allowing to have different "views" of the code, using the basic blocks simply as a representation of the underlying model (the disassembled instructions), so that different basic blocks would contain the same instructions and those basic blocks would accurately represent the flow in the two functions...
In the next figure, the colored basic blocks contain the same instructions in both functions, but the flow is different because of the branching.



I'm leaning towards the second approach (the one in the previous figure), our SQL schema should support it trivially, which is fairly neat.