emit DW_AT_main_subprogram

[tromey]

When making an executable, rustc should emit DW_AT_main_subprogram on the CU and the subprogram DIE for the main function that is going to be executed by main.

This marker would allow gdb's start command to find the correct spot in user code to set an initial breakpoint. (This doesn't work yet, see https://sourceware.org/bugzilla/show_bug.cgi?id=16264 - but could be made to work.)

[michaelwoerister]

I can't find any way to make LLVM do this :/

[tromey]

FWIW I have a patch to LLVM to add the necessary bits. Now I'm investigating how to get the information from the driver to create_function_debug_context.

[michaelwoerister]

See

rust/src/librustc_trans/base.rs

Line 1144 in 43006fc

let (main_def_id, span) = match *ccx.sess().entry_fn.borrow() {

This is how you can get the DefId of the main function.

[tromey]

Ok, I have this working. I sent a patch to llvm-commits, but it needs moderator approval to go through, and then of course patch review.

I also have a rustc patch that works, but I haven't written a test case yet.

One possible issue is that the patch that applies to Rust's llvm is different from the patch I sent upstream. My current plan is to wait until something goes in upstream, then backport it and submit a PR to the Rust fork.

[michaelwoerister]

Sounds like a good course of action :)

[tromey]

http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20161121/407070.html

[tromey]

You can see my patch here: https://github.com/tromey/rust/tree/main-subprogram

I think I can't really submit it until the rust-llvm patch is reviewed, because the final rust patch will have to include a change to the submodule revision.

[tromey]

Also, FYI, the gdb patch is still pending; I will ping it either tomorrow or Friday.

[cuviper]

What happens if you're using an external (unpatched) LLVM? Will it ignore the unrecognized flag?

[tromey]

What happens if you're using an external (unpatched) LLVM? Will it ignore the unrecognized flag?

I think it depends on the precise version of LLVM. The version that Rust is using doesn't check flags less than 1<<27 (IIRC), but LLVM trunk checks that the flag is in range -- which is one reason the upstream LLVM patch is a bit different. I didn't look to see when this change was made in LLVM, but I could if you like.