My name is Artem Vopilov, I am a student at TU Darmstadt. I am writing to you to ask about Alias Analysis.
I am using llvm to analyze alias between variables in programs. I am using Alias Analysis implemented in llvm with command "opt -analyze -aa-eval -print-all-alias-modref-info" and for printing sets of alias "opt -analyze -aa-eval -print-alias-sets".
I execute these commands on the small test program I attached to this email.
If you take a look at it, it will be obvious, that variables "ptra" and "ptrb" alias. However, the results of Alias Analysis do not indicate, that these variable alias. I would like to ask you, if it is possible to change the source code of Alias Analysis, so that it will show that "ptra" and "ptrb" variables from the test program alias.
I am looking forward to hearing from you and hope, that you can help me solve my problem.
The variables ptra and ptrb don’t actually alias. A simple way to think of aliasing is to to say two l-values alias if assigning into one changes the other. In your program changing the value of ptra does not change the value of ptrb. What you do have is two pointers that point to the same l-value. The aliasing analysis will show that variable a (which is address taken) is aliased to a shadow of type int. Any indirections through ptra or ptrb will use that same shadow. This is how clang and llvm get the aliasing correct.
This is a common mistake when learning about aliasing. People often think the pointer should alias the object. It’s really the shadow (i.e. a pointer indirection) that aliases the object. In fact you can never have two variables ever alasing each other.
To Artem:
Next time It will be great if you also attach the IR code you feed into your opt analyze command, the output result, and point out that why you think ptra and ptrb don’t alias.
And basically I agree what Sean stated in this thread: a pointer is just a variable containing memory address, so pointer itself will never alias with other pointers, but the ‘pointee’ will alias with other memory addresses.
I think you two might talk about different things here. As Bekket noted,
include the IR you run!
If I take your test (c code), compile it to IR, and then run mem2reg,
there are _no_ pointers left that could alias. (So you might want to
make the example more meaningful).
If I do not run mem2reg, the two stack locations that are generated for
ptrA and ptrB do not alias as they are _different stack locations_. The
values stored at the stack locations do however alias.
Thank you all for your responses!
I think, they helped me to understand the Alias Analysis. The main thing i understood is that alias analysis will not indicate that variables alias if changing the value of one variable will not affect value of another variable.
But if I want to find all pointers, which point to the same memory location, can I do this with llvm without transforming test program code (with -analyze flag)?
