[NativeAOT] Broken COMDAT support in ObjWriter

[filipnavara]

There are multiple bugs in ObjWriter emission of COMDAT sections:

• For COFF there's always one .debug$S section which is incorrect. For methods emitted in COMDAT sections there should a new associated COMDAT .debug$S section for each (eg. EmitCVDebugFunctionInfo should have similar section handling code to EmitWinFrameInfo).
• For COFF the xdata section with UNWIND_INFO structures is not emitted as associated COMDAT. This mostly works due to unique symbol prefixes but it's not ideal.
• For ELF incorrect relocation types are produced between sections (RELPTR32 is not expressed as PLT relocation). This results in error in ObjWriter if one tries to use ILCompiler with --multifile switch.
• For MachO the support is completely missing. It can likely be implemented with the S_COALESCE feature which serves the same purpose.

[jkotas]

@filipnavara Thank you for finding and reporting these issues. Do you plan to submit fixes for these issues?

[filipnavara]

This particular set of issues is very low priority for me. I don't think the --multifile scenario is properly tested at the moment and I doubt it works outside of very narrow set of conditions. I do plan to submit a batch of ObjWriter fixes which may help a bit as a side effect.

Depending on feedback to #77178 I may reprioritise. Most of the bugs were found when comparing my C# ObjWriter to the output of the LLVM ObjWriter, or when trying to use LLVM ObjWriter with non-default options.

[MichalStrehovsky]

I don't think the --multifile scenario is properly tested at the moment

We run it in the CI with a smoke test so that it doesn't completely bitrot but it hasn't been a big priority:

https://github.com/dotnet/runtime/tree/main/src/tests/nativeaot/SmokeTests/MultiModule

IIRC it never worked outside Windows. These ObjWriter issues were likely the reason.

Multifile needs a bit more work to productize and it's unclear we need it right now (it speeds up compilation, but suppresses a lot of whole program optimizations, which makes it only suitable for debug builds and we have a JIT-based solution for debug builds/inner loop in .NET).

[filipnavara]

We run it in the CI with a smoke test so that it doesn't completely bitrot

The tests run in Release mode. I am not even sure whether they are compiled with debug symbols. If the -g switch was not used, then at least the debug symbol and relocation issues would be untested.

[filipnavara]

The description of the Mach-O mechanism in the original post was inaccurate. Apparently the magic sauce are weak def symbols, which used to have the requirement for S_COALESCED section. LLVM stopped emitting the special S_COALESCED sections though since newer linker doesn't need them. Weak def symbols can be present anywhere.

/*
 * The N_WEAK_DEF bit of the n_desc field indicates to the static and dynamic
 * linkers that the symbol definition is weak, allowing a non-weak symbol to
 * also be used which causes the weak definition to be discared.  Currently this
 * is only supported for symbols in coalesed sections.
 */
#define N_WEAK_DEF	0x0080 /* coalesed symbol is a weak definition */

[MichalStrehovsky]

@filipnavara off the top of your head, do you know if we ported these bugs over to the managed object writer?

[filipnavara]

The COFF bugs were fixed. Mach-O support is still missing. ELF would need to be revisited and rechecked.

[MichalStrehovsky]

Thanks! Will keep this open then.