[lld] Workaround incorrect gcc -ggdb3 .debug_macros generation like mold just did

[Romain-Geissler-1A]

Hi,

This issue relates to some others I had opened in the past, for a problem that seems to be in the end some incorrect .debug_macros generation by gcc when building with gcc -ggdb3.

See the historic of this issue in different projects:

• https://bugs.llvm.org/show_bug.cgi?id=42030
• which lead to this gcc bug: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91239
• and recently mold hit exactly the same issue: Incorrect link of .debug_macro with mold 1.1.1 rui314/mold#438

@rui314 The author of mold decided to eventually workaround this in mold, depsite gcc having some debatable/wrong behavior, in this commit: rui314/mold@8c5e4df

Would it be possible to implement a similar workaround in lld so that binaries built with gcc -ggdb3 and linked with ld.lld don't end up in a huge memory/cpu hog in gdb ?

Cheers,
Romain

[llvmbot]

@llvm/issue-subscribers-lld-elf

[ericriff]

Hi all,

I just came across this myself. I'm wondering if you have any plans to implement a workaround for this like mold did.

Thanks.

Eric.

[rui314]

By the way, did someone file a bug to gcc so that gcc won't create a bad comdat section for -ggdb3? If so, please paste the bug's URL for cross reference.

[Romain-Geissler-1A]

Yes I have opened this gcc bug several years ago, the link is actually in the first message of this llvm issue ;)