Allow preserving name section without other debug info

[trybka]

Usecase:
The name section is similar to the symbol table for ELF binaries. It can be used to symbolize functions at runtime (or after the fact as well).

This is exposed generally via emscripten_pc_get_function library call (and used by the sanitizer stacktraces).
It uses the wasmOffsetConverter which in turn uses the WebAssembly name section.

It is useful to be able to do symbolization without all the other debug info, and for ELF binaries, this is typically possible as long as the binary is not entirely stripped. This is typically the difference between --strip-debug and --strip-all options.

Unfortunately, binaryen and wasm-ld both treat the name section as debug info, and so it is stripped with the rest of the debug information.

I propose that the name section be treated like the symbol table in ELF, and preserved when only discarding debug info, but can be dropped if --strip-all/--strip-unneeded is passed.

[trybka]

+cc @walkingeyerobot

[jeffRTC]

@trybka

but can be dropped if --strip-all/--strip-unneeded is passed

How about just having a new option like --prevent-striping-name-section instead just making it default?

[trybka]

I don't have strong feelings. My preference was for the section to be treated like the symbol table.

Keep if --strip-debug
Discard if --strip-all

But a separate flag is fine, too, considering that my workaround is to use objcopy to drop just the .debug_* sections.

[kripken]

I think this should work today with --profiling-funcs.

[trybka]

So --profiling-funcs does work. Sort of. Inlining seems to be affected which results in a more sparse stacktrace.
This may be because -g was inhibiting inlining in Binaryen.

@dschuff also has out https://reviews.llvm.org/D73820 which seems relevant.

In a case where I'm already setting -g, it still seems like it might be nice to have the name section treated like the symbol table and not discarded if I'm not also passing --profiling-funcs.

[dschuff]

Is it the case that using -g at link time is stripping the name section now? I'm not sure that's expected.

[trybka]

I think what I'm trying to say is that -Wl,--strip-debug at link time discards the name section.

[dschuff]

Yeah, that is probably a holdover from before DWARF even existed. It probably makes sense to treat the name section like ELF symbol tables, and have --strip-debug keep them and -strip-all drop them.

[dschuff]

Looking again at https://reviews.llvm.org/D73820 reminded me that wasm of course also does have a symbol table that plays the same role in linking as ELF's. The difference is that in ELF that linking symbol table is also used for symbolization, and in wasm it's the name section instead. So I think you could make the case for either behavior: 1) the name section plays no role in linking and it's really just for debugging (and profiling) purposes, therefore it makes sense to strip it with --strip-debug. or 2) the name section is used for some of the same purposes as the ELF symbol section (and there are use cases for having it without debug info) and it makes sense to keep when using --strip-debug.

Would changing the behavior of --strip-debug in the linker change any use cases that aren't already using -Wl flags? I think that for llvm-strip, making -strip-debug keep the name section would be ok. Most users of strip just use strip-all (the default), and without this behavior, strip-all is almost exactly the same as strip-debug for linked object files anyway (so making them be actually different seems useful).

[trybka]

Would changing the behavior of --strip-debug in the linker change any use cases that aren't already using -Wl flags?

I'm not sure I understand the question. AFAICT it would be a behavior change, and could be noticed by folks. (i.e. now their binaries are bigger). At least llvm-strip was adding new functionality, rather than behavior changing.

Ultimately, I agree--I can see arguments for either (1) or (2).
One argument for (2) is to keep the linker and strip commands consistent. I think given the changes made to llvm-strip I would be OK if you closed this. Up to you.

[sbc100]

I'm fine with (2) and I agree we should probably keep the meaning of --strip-debug the same between the linker and the strip tool.

I can't imagine there will be many user of wasm-ld --strip-debug who will be upset to learn that the name section is now preserved in this mode in that mode.

Also, the name section is now also used during objdump -d to provide dummy symbols so we are kind of using the way ELF does in this case. We should probably extend that so that llvm-nm also uses them.

[dschuff]

Sounds good to me, let's go with 2).
So that means we should update llvm-objcopy/strip and also lld.

[dschuff]

OK, so it turns out that as of llvm/llvm-project@7cb25f5 llvm-strip -strip-debug actually does preserve the name section, and --strip-all (the default behavior of llvm-strip) removes the name section. So I guess we just need to update wasm-ld

[dschuff]

/cc @sunfishcode in case you have opinions, since this seems nice for non-emscripten use too.