[LLVMdev] Source-level dataflow analysis with LLVM IR

[Saheel Godhane]

Hello everyone!

Greetings! I am Saheel, a PhD student in UC Davis. I have been using LLVM
and Clang for a month now, aiming to do some program analysis. I am sort-of
stuck and am hoping to find some help from the experienced folks here.

So, I am trying to do an intra-procedural dataflow analysis on the lines of
"given a variable declaration return its various definitions, and for each
definition, return the different uses".

After some reading, I thought (correctly?) that if I am able to answer the
question "is a given variable definition live at a given location in
program" (Liveness analysis) my task is pretty much done.

For this, I started out with writing an LLVM FunctionPass (based on def-use
section
<http://llvm.org/docs/ProgrammersManual.html#iterating-over-def-use-use-def-chains>
of the manual and some old LiveValues code
<http://www.opensource.apple.com/source/llvmCore/llvmCore-2358.3/include/llvm/Analysis/LiveValues.h>
I found online). But soon I figured out that LLVM IR generates new
variables (%5, %6, etc.) whenever a program variable is used (i.e. on each
variable load) and thus doing an analysis on the IR would not really answer
my question. Is this correct (which would mean I should work with Clang)?
Or is there a way to do work with LLVM IR and still answer source-level
questions?

I do know that we can track
<http://llvm.org/docs/SourceLevelDebugging.html#c-c-source-file-information>
the IR back to source but I am not sure if the LLVM-generated IR-level
variables can still allow for an analysis on source-level variables.

Any pointers to existing (LLVM or Clang) analyses or help on how to do one
myself would be greatly appreciated! :)

Sincerely,
Saheel Godhane.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20141204/e5bd6671/attachment.html>