llvm dev - Feb 2021 - Some questions about lld with gdb-index option

On 2021-02-27, Sterling Augustine wrote:
>It's very common for some input files to have gnu-pubnames, and some
not.
>So the warning would have to be somewhat smart.
I agree with Sterling that this can be very common:

* system libraries compiled with -g but not -ggnu-pubnames
* third-party libraries compiled with -g but not -ggnu-pubnames

>On Sat, Feb 27, 2021 at 2:48 PM David Blaikie via llvm-dev <
>llvm-dev at lists.llvm.org> wrote:
>
>> I think the gdb-index support in lld only works with debug_gnu_pubnames
in
>> the input files - so you would need to compile with -ggnu-pubnames.
>>
>> Maybe lld could/should have a warning or something if the input files
have
>> debug_* sections but don't have debug_gnu_pubnames when
-Wl,--gdb-index is
>> specified.
The issue is https://bugs.llvm.org/show_bug.cgi?id=34820
We were hesistant on whether ld.lld should learn more DWARF stuff.

The problem is not that serious because it (`p &something` => $address,
no debug
info`) only affects objfiles which haven't been parsed by gdb (something
like a
lazy mode in gdb).  Once an objfile is loaded, everything works fine.



Below are something not related to the topic:
>> On Sat, Feb 27, 2021 at 9:52 AM 何文婷 via llvm-dev <llvm-dev at
lists.llvm.org>
>> wrote:
>>
>>> Hi all,
>>>     I am using below command to generate my object file named
`myelf`:
>>>     `clang++-10 -Wl,--dynamic-linker,/lib64/ld-linux-x86-64.so.2
-fuse-ld=/.../usr/bin/ld.lld -rdynamic -Wl,—gdb-index -o myelf xxx.a xxx.a
xxx.a`
-fuse-ld=/absolute/path or -fuse-ld=relative/path is not recommended.
You'll get a warning with -Wextra.

(
The reason is that many toolchains do hacky find(path, "lld") style
string match to know whether the linker is ld.lld and do something magic.
)

Use --ld-path=/path/to/ld.lld

Linking a bunch of .a files without an .o may not do what you want.
If .a does not defined a symbol which is previously undefined, the .a file will
be dropped from the link.

https://sourceware.org/binutils/docs/ld/Options.html "-l namespec" and
https://lld.llvm.org/ELF/warn_backrefs.html
have some information.
>>>     I found when link with -gdb-index option, it will generate the
section of .gdb_index in my elf file named myelf. But this gdb_index section is
not full, and when I gdb myelf to print some function like `abc` , it shows that
no symbol found(.gdb_index section donot have the function abc, but .debug_full
and .symtab has this function).
The two-dash form --gdb-index is recommended.
>>>     But When  I am using clang++ compile without gdb-index option 
and then  using `gdb-add-index myelf `command to add gdb-index section, this
section is much larger, and When I gdb to print some function, all the symbols
can be found.
>>>     I do not understand, why lld generate a smaller .gdb_index
section, and is there any options to let me generate full .gdb_index section?
>>>
>>>     Waiting for some advices.
>>>
>>> Best wishes.
>>> hexiaoting
>>> 2020.2.25
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> LLVM Developers mailing list
>>> llvm-dev at lists.llvm.org
>>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>>
>> _______________________________________________
>> LLVM Developers mailing list
>> llvm-dev at lists.llvm.org
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>>

On Sat, Feb 27, 2021 at 6:20 PM Fangrui Song <maskray at google.com>
wrote:
> On 2021-02-27, Sterling Augustine wrote:
> >It's very common for some input files to have gnu-pubnames, and
some not.
> >So the warning would have to be somewhat smart.
>
> I agree with Sterling that this can be very common:
>
> * system libraries compiled with -g but not -ggnu-pubnames
> * third-party libraries compiled with -g but not -ggnu-pubnames
>
Sure, but the resultant broken gdb behavior is pretty problematic to
experience without some notification.

If gdb-index had a way to say which CUs it covers (.debug_names has this, I
believe), that'd be great - but without that, I think it's fundamentally
invalid/incorrect to produce an incomplete gdb-index.

> >On Sat, Feb 27, 2021 at 2:48 PM David Blaikie via llvm-dev <
> >llvm-dev at lists.llvm.org> wrote:
> >
> >> I think the gdb-index support in lld only works with
debug_gnu_pubnames
> in
> >> the input files - so you would need to compile with
-ggnu-pubnames.
> >>
> >> Maybe lld could/should have a warning or something if the input
files
> have
> >> debug_* sections but don't have debug_gnu_pubnames when
-Wl,--gdb-index
> is
> >> specified.
>
> The issue is https://bugs.llvm.org/show_bug.cgi?id=34820
> We were hesistant on whether ld.lld should learn more DWARF stuff.
>
I think it should be: either create a complete gdb-index, or error.
Creating an incomplete gdb-index is/should be considered a fairly severe
bug.

> The problem is not that serious because it (`p &something` =>
$address, no
> debug
> info`) only affects objfiles which haven't been parsed by gdb
(something
> like a
> lazy mode in gdb).  Once an objfile is loaded, everything works fine.
>
I'd say this is pretty serious - unreliable operations lead to lack of
confidence/pretty deep frustration with tools.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210228/5c217fb1/attachment.html>