[LLVMdev] llvm DSA - reproduce the result in PLDI 07 paper

[Wan Zhiyuan]

Dear all,

I am trying to reproduce the "Percent May Alias" result described in PLDI 07's paper "Making Context-Sensitive Points-to Analysis with Heap Cloning Practical For The Real World" (http://llvm.org/pubs/2007-06-10-PLDI-DSA.html).

However, my "Percent May Alias" for all the benchmarks is much greater, especially "bzip2".

The DSA code I use is from "https://github.com/smackers/smack". I have diff the code between smack and poolalloc_32. They are almost the same except the "#include" statements.

I was wondering whether I need to do some configuration to make DSA work properly.

Thank you!

Zhiyuan

[John Criswell]

Dear Zhiyuan,

In order to reproduce the results from the paper, you'll need to replicate a system from that era.  You'll need to use the same version of LLVM and DSA that the paper used.  I think that was LLVM 1.9 (the release_19 branch of LLVM and poolalloc), but I'm not sure.  You should check to see if the paper specifies the version.

As you'll be using a very old version of LLVM, it may be worth setting up a VM with a corresponding old version of Linux.  I suspect newer compilers will not compile a version of LLVM that is that old, so using an old version of Linux with an old version of GCC may be needed.  I think Fedora Core 2 is the OS we were using at the time.

To answer the question of why you can't use a modern version of LLVM and poolalloc, it's because LLVM has changed significantly.  DSA relies upon the type annotations provided in the LLVM IR to "bootstrap" its type inference (bootstrap is not quite the right word, but it's the closest one of which I could think).  As LLVM matured, transformations would ditch the type information (e.g., transforming typed GEPs into untyped GEPs into a byte array), making DSA's ability to do type-inference (and thereby improving field sensitivity) more difficult.  Throw into the mix the fact that DSA is maintained by an academic research group, and the result is that modern DSA doesn't have the accuracy that the original DSA did.

The good news is that I think DSA can be fixed by making its type-inferencing code smarter.  The bad news is that it'd be a fair amount of work to do.  So far, no one has had sufficient desire/motivation to design and implement the improvements.

Regards,

John Criswell

_______________________________________________
LLVM Developers mailing list
LLV...@cs.uiuc.edu         http://llvm.cs.uiuc.edu
http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev

-- 
John Criswell
Assistant Professor
Department of Computer Science, University of Rochester
http://www.cs.rochester.edu/u/criswell

[Wan Zhiyuan]

Dear John,

I intend to implement the improvements on DSA.

After running DSA on SPEC, I found DSA gives low precision for mcf and bzip2.

I have checked the most imprecise c files in mcf an found that the code seems to be a mixture of "PHI" and "GEP" instructions.

Could you please give me some hints about what the big picture of the improvement should be and how to start?

Thank you!

Regards,

Zhiyuan

[John Criswell]

On 5/5/15 5:33 PM, Wan Zhiyuan wrote:

Dear John,

I intend to implement the improvements on DSA.

There be nasty dragons in DSA.  Don't say I didn't warn you.
:)

On a side note, I recently discovered that someone (I think Will Dietz) updated the code to work with LLVM mainline, so you should be able to update to a newer version of LLVM and use DSA.  I've created a snapshot of the LLVM source tree and a modified version of the poolalloc tree in https://github.com/jtcriswell/llvm-dsa that you may find useful.  I think Will is also maintaining a github tree that he updates regularly.

After running DSA on SPEC, I found DSA gives low precision for mcf and bzip2.

I have checked the most imprecise c files in mcf an found that the code seems to be a mixture of "PHI" and "GEP" instructions.

Could you please give me some hints about what the big picture of the improvement should be and how to start?

There may be many reasons why DSA's precision is low.  You'll need to figure out what the reason is for each program.  In some cases, it may be type inference.  In other cases, it may be that DSA is cognizant of the influence of external code (which causes it to correctly give pessimistic results).  For programs using function pointers, there's a bug in the algorithm in which DSNodes that should be marked complete are not.  It could also be that you're using the wrong DSA pass (e.g., using EQTD when TD will do).

Looking at type inference specifically, the problem, in a nutshell, is that DSA's type inference should not rely upon LLVM's type annotations.  It should just create a map from offsets to types.  Some work on this has already been done (e.g., DSA can figure out that casting a pointer to a structure into a pointer to the type of the structure's first field is still offset 0).  However, there are still places in which DSA is relying upon LLVM's type annotations.

One thing that I would like to look at is changing how DSA analyzes arrays of structures.  Right now, DSA tries to infer a structure type for a memory object and then tries to infer whether the memory object is a singleton structure or an array of structures (you can see this in DSA's interface; you can see a map between offsets and types and an 'A' flag indicating that the object is an array).  I think this makes DSA needlessly complicated.

I think it would be better if DSA did what Value Set Analysis (VSA) does.  VSA was designed to analyze untyped binary code.  For each abstract memory object, it creates a map of 4-tuples to types.  Each 4-tuple represents a formula ax+b as (a,b,c,d) in which b if offset, a is stride, and x is constrained between values c and d (c and d can be constants or +-infinity).  For example, if you have an array of struct {int a ; char * b} on a 32-bit machine, DSA currently tries to figure out that there's an array of structure elements in which there's an int at offset 0 and a char * at offset 4 within the structure.  VSA would say that there's a memory object with an int at every multiple of 4x+0 offset and a char * at every 4x+4 offset.

The VSA approach should be agnostic to all sorts of weird casting, embedded arrays, and embedded unions, though this is an educated guess at present.

Regards,

John Criswell

[Wan Zhiyuan]

Dear John,

Thank you so much for your reply!

Please kindly correct me if my understanding is wrong. 

From your reply, I learn that there could be mainly four reasons why DSA's precision is low:

1. type inference

2. influence of external code

3. bug in handling function pointers

4. using the wrong DSA pass

As you mentioned in your previous mail, LLVM changes a lot. 

So I was wondering if DSA's algorithm (presented in PLDI '07) can achieve a comparable precision on top of LLVM 3.5+ after implementing the improvements.

Also, I have diff the DSA code in poolalloc release_19 and release_32.  It seems to me that the changes in DSA mainly comes from the aspects as below:

1. release_32 introduces the DataLayout to do some type inference;

2. release_32 replaces some data structures with LLVM built-in data structures;

3. release_32 handles different types of LLVM IR instructions, and in a different way.

4. the inheritance of Analysis Passes has changed a bit.

Therefore, I was wondering if the introduction of DataLayout and changes in LLVM IR cause the imprecision as well.

Thank you!

Regards,

Zhiyuan Wan

[John Criswell]

On 5/6/15 11:07 PM, Wan Zhiyuan wrote:

Dear John,

Thank you so much for your reply!

Please kindly correct me if my understanding is wrong. 

From your reply, I learn that there could be mainly four reasons why DSA's precision is low:

1. type inference

2. influence of external code

3. bug in handling function pointers

To be clear, this is a fault in the algorithm itself (i.e., a completely correct implementation would still have the problem).  No one at the time was aware of this flaw in the algorithm; a master's student discovered it several years after the paper was published.

4. using the wrong DSA pass

As you mentioned in your previous mail, LLVM changes a lot. 

So I was wondering if DSA's algorithm (presented in PLDI '07) can achieve a comparable precision on top of LLVM 3.5+ after implementing the improvements.

That's an open question and dependent on a lot of things.

I also think you're asking the wrong question.  It sounds like you want to fix DSA because doing so will make it more precise, but you haven't discussed why you need that precision.

In my opinion, fixing DSA will be a lot of work (2-3 months at least).  I would not spend time fixing DSA just to see if some synthetic precision results get better.  I would fix DSA only if I was writing an analysis or transformation that needed points-to, type-inference, and/or call graph results more accurate than what DSA is currently providing *and* if fixing DSA was the cheapest option in getting those more accurate results.

Also, I have diff the DSA code in poolalloc release_19 and release_32.  It seems to me that the changes in DSA mainly comes from the aspects as below:

1. release_32 introduces the DataLayout to do some type inference;

Not really.  DataLayout used to be called TargetData.  Someone changed the name within the past year or so.

2. release_32 replaces some data structures with LLVM built-in data structures;

3. release_32 handles different types of LLVM IR instructions, and in a different way.

4. the inheritance of Analysis Passes has changed a bit.

Therefore, I was wondering if the introduction of DataLayout and changes in LLVM IR cause the imprecision as well.

You're looking at all the cosmetic changes needed to make the DSA code compile with modern LLVM.  None of these things is what makes DSA less precise.

When I say that changes in LLVM have caused DSA to become less precise, I am talking about two issues.  First, newer versions of the LLVM front-end generate LLVM IR that DSA does not analyze well.  For example, after LLVM 1.9, the C front-ends started changing the way they generated code for vararg functions.  DSA treated the new code conservatively.

Second, some of the LLVM optimization passes started replacing well-typed GEP instructions with code that casted the pointer to a char *, did a GEP, and then casted the pointer back to its original type.  Since DSA wasn't designed to analyze this idiom, it just thought that the pointer was being used like an untyped byte array.  This made DSA's results a lot less precise because it dropped field sensitivity for these memory objects.

My suggestion on replacing DSA's type-inference code with something similar to VSA is my idea on how to address the second issue in a way that will avoid having to change DSA every time the LLVM optimizations change what they do.  In essence, by making DSA less reliant on the LLVM type information, the LLVM optimizations can do whatever they want with the type information, and DSA should still continue to get the same results.

To summarize, DSA lost precision because it made assumptions about the LLVM IR it would be analyzing.  When LLVM changed in ways that broke those assumptions, DSA started generating less precise results.