Advice on memory copy instrumentation

Project Infrastructure LLVM Dev List Archives
#1

Hi all,

Given a couple of lines of C++ code int x = 42; int y = x, we end up with the following LLVM IR instructions:

%x = alloca i32, align 4
%y = alloca i32, align 4

store i32 42, i32* %x, align 4
%0 = load i32, i32* %x, align 4
store i32 %0, i32* %y, align 4

Is it possible to instrument the IR to perform a value trace?

What I’d like to do is stream a log of memory copies (reads then writes), such that in this example, the final instruction will produce a log event along the lines of:
"COPY: Value at copied to "?

Essentially what I’d like to do is annotate particular values, so that when these same values are encountered again later in the program, I can retrieve the annotation. I will also need the annotation to survive copies, moves, etc. This could be considered a lightweight, parallel symbolic trace performed at runtime on a very small subset of program values.

I am hoping to implement this tooling at the LLVM IR level, so that it can be useful beyond C++, but if it’s easier to instrument the CLang AST instead then I guess I can start there. Looking forward to your responses :slight_smile:

#2

I’m not sure if I understand exactly what you would like to do, but I would imagine one could write a pass that looks for the pattern of “%x = load … store %x …” and insert a function call that logs the copy. I don’t know of any existing analysis or utility code that you could build on to make it easy to do that, though.

It might be helpful if there is an example of the annotations you have in mind.

#3

Thanks Hiroshi, this is exactly what I was looking for! I’ve been looking into it for a few days, and I’ve just found the MemorySSA (virtual IR) which is probably gonna help me out a bunch. Thanks again :slight_smile:

#4

Hi Matt,

You might want to consider running mem2reg first to eliminate trivial
memory locations+writes+reads. See:
  http://lists.llvm.org/pipermail/llvm-dev/2020-February/139405.html

It is important to do that because otherwise almost no existing LLVM-IR
pass will be effective. If you compile with -O{1,2,3} it happens for
you: https://godbolt.org/z/GJXUiS

Cheers,
  Johannes