AST of whole program

Hello,

I would like to do transformations on AST of a c program but I need to have access to all ASTs created for the program to do right changes. LLVM processes one translation unit at a time and because of it, I do not have access to AST of all the translation units at the same time. Do you have any suggestion how I can access all the ASTs created for a program, do analysis on the ASTs and do modifications on the ASTs?

As a summary:

  1. I need to have access to ASTs of the program at the same time.
  2. Do analysis on ASTs.
  3. Modify ASTs based on my analysis and create llvm IR from modified ASTs.
    Thank you,

Anil

- llvm-dev, + cfe-dev

Hi,

What kind of transformation are you interested in, and what kind of programs
are you looking to transform?

By 'AST of whole program', do you mean AST's for the source from all libraries
linked into the program?

vedant

Hi Vedant,

What kind of transformation are you interested in, and what kind of programs
are you looking to transform?

I am trying to change the layout of fields(randomize) in a struct in a c program. I already figured out how to change the layout of fields in a struct but there are structs that I should not touch like structs defined in libraries. So if I randomize a struct in one compilation unit and then realize that actually, I shouldn’t have randomized it when clang was working on another compilation unit, there is no way to go back and revert the layout of the struct that I already randomized in previous translation unit because it is already over. So what I am thinking is I should look all the translation units in AST level before they create llvm IR and decide which structs I should randomize, then randomize the structs I have decided to randomize, then let clang to create llvm IR using modified ASTs.

I am trying to transform programs like apache.

By ‘AST of whole program’, do you mean AST’s for the source from all libraries
linked into the program?

I am not sure if I understand your question but I will try to explain what I meant. For each translation unit, AST gets created. The problem is I can only see AST of current translation unit. I cannot see AST of next translation unit because clang works on one translation unit at a time. Maybe I should dump AST of each translation unit to the disk, decide which structs can be randomized, change the AST on the disk and start compilation from modified ASTs. But this may be so slow and I do not really know how I can do this.

Also I am not sure about one thing. Can I make sure that a struct is defined in a library or in the source code of the program by looking only one translation unit without any false flag? If I can, then there is no need for what I am asking for.

I hope that I explained what I am trying to do clearly. If you have any suggestion how I should do this, I would really appreciate hearing your opinion. Thank you very much for your quick response.

Anil

I am trying to change the layout of fields(randomize) in a struct in a c program.

Have you taken a look at this paper?

    https://www.utdallas.edu/~zxl111930/file/DIMVA09.pdf

The authors describe their approach to data structure layout randomization in
some detail, including pros/cons of implementing the feature at the AST level.
There are some other interesting bits, like their decision to introduce an
explicit obfuscation struct attribute.

... there are structs that I should not touch like structs defined in libraries.

Having an explicit "reorder" attribute is one way to work around this.

So if I randomize a struct in one compilation unit and then realize that actually, I shouldn't have randomized it when clang was working on another compilation unit, there is no way to go back and revert the layout of the struct that I already randomized in previous translation unit because it is already over.

It sounds like you need some rule that tells you whether or not it's OK to
randomize a struct, regardless of how many translation units you have already
processed.

So what I am thinking is I should look all the translation units in AST level before they create llvm IR and decide which structs I should randomize, then randomize the structs I have decided to randomize, then let clang to create llvm IR using modified ASTs.

Let's say you take this approach, and you collate the AST's for every source
file in a project. What information do you plan on gathering that will help you
determine the right structs to reorder? Can you guarantee that your decision
procedure will never reorder a struct that is not meant to be reordered, and
will always reorder all other structs?

I am trying to transform programs like apache.

I suspect that you'd need to manually audit that codebase and apply "reorder"
attributes to get good results. I could be wrong though :).

Also I am not sure about one thing. Can I make sure that a struct is defined in a library or in the source code of the program by looking only one translation unit without any false flag? If I can, then there is no need for what I am asking for.

In the paper I linked to, the authors mention several other conditions under
which it's inappropriate to randomize structs.

vedant

There are plenty of “not a library” places where re-ordering/changing the data structure will cause trouble. For example storing data in binary form in a file (and plenty of simple applications just write the struct data straight to a file, so order is important) - rebuilding the app, or having multiple applications that are built using the same header.

Applications that share content using shared memory or memory mapped files would be another case where re-ordering the data in any unpredictable way will be disastrous. (I’ve written applications where there are several instances of the actual application, and then a separate status application, where the actual application is storing “what I’m doing right now” in shared memory, and the status application reads a snapshot every second or so, displaying any changes in what’s going on in each of the applications)

So for sure, this sort of thing needs to be done in a way that can either be predictable over two different binaries, and/or be possible to turn off for certain data structures.

Vedant and Mats,

Have you taken a look at this paper?

I looked at that paper and here are the cons the paper points out:

  • If a data structure is used in network communication, the communicating parties may not understand each other if the data structure is randomized.

  • If a data structure definition is public (e.g., defined in shared library stdio.h), it cannot be randomized.

  • There is a special case in GNU C that allows zero-length arrays to be the last element of a structure (a zero-length array is actually the header of a variable-length object). If a zero-length array is declared as the last element in a struct, that element cannot be randomized, otherwise, it cannot pass GCC syntax checking.

  • A programmer may directly use the data offset to access some fields. (This is particularly true in programs which mix assembly and C code.)

  • To initialize the value of a structure, the programmer uses the order declared to initialize the structure. These fields cannot be randomized, as the program may crash.

  • The programmer needs to decide which data structure is randomizable in the program by inserting new keyword “obfuscate”.

Actually, there is another paper that solves some of the cons. Here is the link: http://link.springer.com/chapter/10.1007%2F978-3-642-18178-8_16
In this paper, three of the cons explained above solved. They are as follows:

  • To initialize the value of a structure, the programmer uses the order declared to initialize the structure. These fields cannot be randomized, as the program may crash.

  • There is a special case in GNU C that allows zero-length arrays to be the last element of a structure (a zero-length array is actually the header of a variable-length object). If a zero-length array is declared as the last element in a struct, that element cannot be randomized, otherwise, it cannot pass GCC syntax checking. → solved by keeping zero-length array in the same position.

  • The programmer needs to decide which data structure is randomizable in the program by inserting a new keyword.

There is also another paper that solves some of the other cons by doing randomization dynamically.

My aim is for now implementing frontend plugin or something similar to handle these three cons in the clang.

Having an explicit “reorder” attribute is one way to work around this.

I did not know that there is such a thing. I googled and found out that it is in GCC and it is not supported anymore. Even it is supported I cannot use it because I should not modify the source code myself and I need to do this in clang. I couldn’t find any info whether it is supported by clang and if you have any info, I would be happy to hear that.

It sounds like you need some rule that tells you whether or not it’s OK to
randomize a struct, regardless of how many translation units you have already
processed.

My plan is as follows:

  1. To understand whether a function is library function or not, I will check if the function has only declaration or declaration + definition.
  2. If the function has only declaration, in other words; it is a library function, I will not randomize structs that are passed as parameters or if it returns a pointer to a struct.
  3. I will randomize the rest by handling zero-length arrays and also initialization of struct by values.

I am just not sure whether my plan will always work to eliminate the 3 cons if I do randomization like above in one translation unit at a time, That’s why I wanted to see ASTs of all translation units at the same time. What do you think about it? Do you have any example that will show it will not work if I do it in one translation unit at a time? If you have any example, what is your suggestion, how should I proceed?

I did this randomization as an llvm pass but only randomized the same size fields with each other. If I do randomize all the fields with each other, the program crashes because of align info in store and load instructions.

Anil

Hi Vedant,

*What kind of transformation are you interested in, and what kind of
programsare you looking to transform?*

I am trying to change the layout of fields(randomize) in a struct in a
c program. I already figured out how to change the layout of fields in a
struct but there are structs that I should not touch like structs defined
in libraries. So if I randomize a struct in one compilation unit and then
realize that actually, I shouldn't have randomized it when clang was
working on another compilation unit, there is no way to go back and revert
the layout of the struct that I already randomized in previous translation
unit because it is already over. So what I am thinking is I should look all
the translation units in AST level before they create llvm IR and decide
which structs I should randomize, then randomize the structs I have decided
to randomize, then let clang to create llvm IR using modified ASTs.

I am trying to transform programs like apache.

This problem falls within a category of related problems, where you want
cross-source-file analysis in order to determine how to make
per-source-file changes. The way we usually solve those kinds of problems
is with a two-pass system: the first pass processes one translation unit at
a time, and performs a local analysis to determine what to do under what
circumstances. The local analysis results are then merged to determine what
transformations to perform. Then a second pass over all translation units
performs the actual code changes.

In your specific case, perhaps you could form a list for each TU of the
functions defined locally, and the functions to which each
potentially-reorderable type is passed. Then merge to form a complete list
of function definitions controlled by the program, and determine whether a
type is non-reorderable based on whether it's passed to a function that's
not in your list.

In addition to not requiring a whole-program AST, this approach also
parallelizes naturally (see http://www.hyrumwright.org/papers/icsm2013.pdf
for more on how we use this kind of approach at Google).

Hello Richard Smith,

This makes a lot of sense. This is definitely the approach I will follow. Thank you!

If you know any source that you think it will help me to accomplish two-pass system efficiently other than clang source code, I will be happy to hear that. Thanks again!

Anil