Invalid debug offset invariant

[guybedford]

Replication instructions in rustwasm/wasm-bindgen#3483 (comment).

The last instruction address from https://github.com/rustwasm/walrus/blob/main/src/module/debug/dwarf.rs#L214 ends up being after the subsequent function end address in https://github.com/rustwasm/walrus/blob/main/src/module/debug/dwarf.rs#L210.

Basic logging provides:

Instruction offset: 4834
Instruction offset: 4852
Instruction offset: 4859
Instruction offset: 4869
Instruction offset: 4884
Instruction offset: 4888
Instruction offset: 4898
Instruction offset: 4908
Instruction offset: 4909
Address conversion: 10337 -> 4407 (- 365 = 4042)
Function end: 4042

Where the broken invariant is then that 4909 < 4042 does not hold.

I tested the original PR and it also throws the exact same assertion so this was not an issue with the rebasing.

We maintain separate offset tables for the functions and instructions, and in this case, the instruction offsets are not present so we are falling back to the new None path in the convert_address function (which was not a part of the original Gimli implementation this came from that instead errored).

Separately, it's worth noting there's a language translation issue in the implementation where Instrument is used instead of Instruction, which might help code understanding to know.

[Stefan-Hanke]

I'd like to tackle this - but beware I have not done system-level programming for quite some time. My main motivation is gaining experience in the rust ecosystem in general.

What I did so far:

• Replicated the issue locally - I was glad that the code-lldb VSCode extension worked without any other configuration on Windows, and that I found about that patch table in config.toml (neat stuff)
• Skimmed over DWARF v5 - never been that far down.
• ... yeah and then trying to understand line programs. The spec is a difficult read for me. I found some resources but yeah, need to bite through that.

I'll now go on with generating some ELF debug-built object files (I failed to cross-compile a hello world...) and then using dwex in the hope of gaining insight.

I'd appreciate resources about line programs.

[guybedford]

@Stefan-Hanke good luck! As I say I suspect the issue is somewhere in the offset handling code landed in the original PR that added this work, the None path in convert_address did seem a bit odd to me. If you have any questions at all in looking into this, happy to help in any way I can.

[d3lm]

Is there any progress on this? I really think that having DWARF symbols for wasm-bindgen would be a really good addition an would drastically help with debugging.

[squillace]

Is there anything we can do to help finish this off here? I'd love to have rust support in vscode: rustwasm/wasm-bindgen#2389 (comment)

[guybedford]

@squillace I'll gladly review any contribution to a bug fix. This issue should still have the required context to investigate further.