Different iterator instructions reported in seemingly identical loops

[robcasloz]

Hi Chris, thanks for the effort you put maintaining this library :)

In a simple single-loop function, the iterator recognition pass reports different LLVM IR instructions as belonging to the loop's iterator depending on whether the function returns:

void foo() {
  int i;
  for (i = 0; i < N; i++) {}
  return;
}

or exits:

void foo() {
  int i;
  for (i = 0; i < N; i++) {}
  exit(0);
}

The loops are identical in both cases, so I would expect the same instructions to be reported.

How to reproduce

1. Run clang -emit-llvm -S foo.c -o foo.ll (foo.c is available here: foo.c.zip).
2. Run opt -itr-annotate -S foo.ll -o foo-itr.ll.
3. Assert that grep "iterator \!2" foo-itr.ll and grep "iterator \!3" foo-itr.ll produce the same output. Currently, this step fails.

Thank you for your time! Let me know if I can help in any way.

[compor]

Hey!

This is a quick response, but I'll look into it in depth later today.

So, without having a look at the attached files, I'd say that the IR generated is different because of exit(), which is known to terminate the program wherever/whenever it is called. In IR terms we'd get an unreachable terminator instruction. I'm not sure how that affects the analysis without a further look.

[compor]

A short update: It seems that the control dependence graph (CDG) generated for bar() (which uses exit()) is empty, which most likely means that the specific terminator instruction type (i.e. unreachable) is not handled.

[compor]

@robcasloz This was an issue with Pedigree, so I opened an issue for it here and it is hopefully address in the referenced PR.

Please try it and let me know.

[robcasloz]

Hi @compor, I tried the PR and it works just fine, now the same instructions are reported as iterators in both cases. Thanks!

[compor]

Cool! I'm closing this since it was addressed here.
Thanks for catching this and all the feedback! 😄