SSA for memory objects

Hello,

I'm looking to learn more about what is available for memory versioning or inducing some kind of SSA for memory objects. I found this page:

   http://llvm.org/docs/tutorial/OCamlLangImpl7.html#memory-in-llvm

Which says: "[LLVM] does not require (or permit) memory objects to be in SSA form.... In LLVM, instead of encoding dataflow analysis of memory into the LLVM IR, it is handled with Analysis Passes which are computed on demand."

However, there is no mention of which analyses provide this. I also haven't been able to turn up much in the way of documentation or examples, so I'm grateful for any additional pointers.

Thanks,

Tyler

Dear Tyler,

First, you *could* represent memory objects in SSA form now that LLVM supports vector and structure-based SSA virtual registers. However, there is no LLVM pass that will take heap-based memory objects and convert them into SSA form. Moreover, the code generators are not expecting the compiler to put heap variables into SSA form, and so if you can generate native code from a program that has been transformed in this way, it will probably not run very fast.

Second, regarding analysis passes, I'm not sure that LLVM really has passes for doing data-flow analysis on values in memory. The best I can think of is the DSA points-to analysis; this lives in the poolalloc source code and isn't used by LLVM itself. It has been used by a few research projects to do data-flow analysis on values traveling through the heap.

There are other analyses within LLVM that may reason about heap objects using the AliasAnalysis interface (the passes to remove redundant loads and stores comes to mind), but I don't think they do proper Kam-Ullman data-flow analysis on data traveling *through* heap objects (but someone can correct me if I'm wrong on that).

Regards,

John Criswell

Hi John,

Thanks for your reply.

Quoting John Criswell <jtcriswel@gmail.com>:

First, you *could* represent memory objects in SSA form now that LLVM supports vector and structure-based SSA virtual registers.

Could you elaborate a bit more on what this means? The performance concern isn't a big issue for me at this point, as I'm currently just learning more about data flow analysis on memory objects (i.e. how tractable it is. SSA for memory was just the first idea I had).

Second, regarding analysis passes, I'm not sure that LLVM really has passes for doing data-flow analysis on values in memory. The best I can think of is the DSA points-to analysis; this lives in the poolalloc source code and isn't used by LLVM itself. It has been used by a few research projects to do data-flow analysis on values traveling through the heap.

I'll have a look at DSA. In the meantime, do you happen to have any pointers to those research projects you mentioned?

There are other analyses within LLVM that may reason about heap objects using the AliasAnalysis interface (the passes to remove redundant loads and stores comes to mind), but I don't think they do proper Kam-Ullman data-flow analysis on data traveling *through* heap objects (but someone can correct me if I'm wrong on that).

My last resort was going to be rolling my own analysis using AA, but I first wanted to see what work has been done in this area so I can avoid reinventing the wheel.

Thanks,

Tyler