Modify context generation slightly (xrefs)

[trashvisor]

Truncate xrefs, so that the source code of the entire function(s) xref'ing the function under test is not included.

Additionally, modify context generation so we talk about lines where the code is xref'd, as opposed to functions.

[DonggeLiu]

nit: lines[1:2] will not include line index 2.

start = max(0, line_index - 10) 
end = min(len(lines), line_index+10)

[DonggeLiu]

/gcbrun request_pr_exp.py -n dg-272

[DonggeLiu]

@trashvisor Could you please check if the report:

1. Has any unexpected behavior (e.g., failed to include certain xrefs)
2. Significant improvements / regressions?
   https://llm-exp.oss-fuzz.com/Result-reports/ofg-pr/2024-05-15-272-dg-272-comparison/index.html
   Thanks!

if not, I reckon this is ready to merge.

[oliverchang]

/gcbrun request_pr_exp.py -n oc-272

[DonggeLiu]

A silly question:
What is the intention of this block in the prompt?
Is it to mitigate the 'undefined function' error?

I noticed that it adds the function declaration, which is the same as the function signature added earlier.

Related:

oss-fuzz-gen/llm_toolkit/prompt_builder.py

Line 165 in 39bebba

must_insert=context_info['decl'],

[trashvisor]

A silly question: What is the intention of this block in the prompt? Is it to mitigate the 'undefined function' error?

I noticed that it adds the function declaration, which is the same as the function signature added earlier.

Related:

oss-fuzz-gen/llm_toolkit/prompt_builder.py

Line 165 in 39bebba

must_insert=context_info['decl'],

It was originally intended to solve:

• Not having the appropriate header file for the function declaration (so we would just embed the function decl itself without a header and things might work).
• Compatibility with c programs, as an extern "C" would be added to solve linkage issues.

Should this block be removed or some templating logic added?

[DonggeLiu]

It was originally intended to solve:

• Not having the appropriate header file for the function declaration (so we would just embed the function decl itself without a header and things might work).

Ah, thanks for the explanation!

I doubt this will help LLM get the correct header file, though:

1. We already have the signature in the prompt, so this doesn't add new information.
2. LLM is not good at searching for the corresponding header file given the function signature, it can only guess.

But don't worry; AdaLogic will look into finding the correct header file in a more programmatic way.

• Compatibility with c programs, as an extern "C" would be added to solve linkage issues.

In this case, maybe:

1. Make the instruction more explicit, e.g.,

Include the following extern C linkage specification in the fuzz target, because the function-under-test is defined in C and the fuzz target is written in C++:
extern ''C" <function-signature>

2. Only add this code if the project-under-test is in C.

Would this sound reasonable? Thanks : )