Segfaults when builder is called without setting its position first

[jdupak]

Describe the Bug
Program compiles successfully but segfaults, if builder position is not set to some BB.

To Reproduce

        let main = module.add_function("main", ctx.i64_type().fn_type(&[], false), None);
        let entry = ctx.append_basic_block(main, "entry");
        // builder.position_at_end(entry); without this is segfaults
        let exit_code = builder.build_alloca(ctx.i64_type(), "exit_code");
        let exit_bb = ctx.append_basic_block(main, "exit");

Expected Behavior
This should at least produce a "controled" runtime error.

LLVM Version (please complete the following information):

• LLVM Version: 11, 14
• Inkwell Branch Used: crates.io release 0.2.0

Desktop (please complete the following information):

• OS: Manjaro

[TheDan64]

I wanna say this is a known issue but I can't find a ticket for it lol.

Not sure the LLVM api has a safe way for us to verify if a position has been set

[marcospb19]

I was about to open an issue but found this one.

Minimal reprod:

fn main() {
    let context = inkwell::context::Context::create();
    let builder = context.create_builder();

    // [1] 70215 segmentation fault (core dumped)
    builder.build_alloca(context.i64_type(), &"n");
}

inkwell = { version = "0.2.0", features = ["llvm15-0"] }

[EricLBuehler]

if builder position is not set to some BB.

Perhaps Inkwell can store a state of BB, so that it has a runtime error? Or better yet, a type-state of some sort?

[TheDan64]

Yeah, I've thought about using type state for this before but we'd have to figure out which methods are valid for each state and there are a lot of builder methods

[EricLBuehler]

I agree, it would be complicated to implement type state.

To implement type state based safety, perhaps a builer pattern can be implemented for the Builder struct. With this idea, the only way a "buildable" Builder can be created is by calling one of the position_* methods on a "non-buildable" Builder? This would return a "buildable" Builder and would essentially be a build method for the builder pattern.

However, as someone who uses Inkwell in my own projects, I think this would be cumbersome from an API useability point of view. For example, a Builder<State=NotSet> and then an Option<Builder<State=Set>> would need to be kept in order to use type state safety. In addition, backwards compatibility would be an issue. However, the major improvement this would bring is that no segfaults are caused by a simple mistake.

For those reasons, I am think runtime assertions to ensure safety by adding state to the Builder struct would be better.