[PATCH v1 1/1] bfd: elflink: upstream change to garbage collection sweep causes mb regression

[PATCH v1 1/1] bfd: elflink: upstream change to garbage collection sweep causes mb regression

[Neal Frager]

Upstream change for PR13177 now clears the def_regular during gc_sweep of a
section. (All other archs in binutils/bfd/elf32-*.c received an update
to a warning about unresolvable relocations - this warning is not present
in binutils/bfd/elf32-microblaze.c, but this warning check would not
prevent the error being seen).

The visible issue with this change is when running a c++ application
in Petalinux which links libstdc++.so for exception handling it segfaults
on execution.

This does not occur if static linking libstdc++.a, so its during the
relocations for a shared lib with garbage collection this occurs.

Signed-off-by: Neal Frager <neal.frager@amd.com>
---
 bfd/elflink.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/bfd/elflink.c b/bfd/elflink.c
index ca162145f7e..0524019641e 100644
--- a/bfd/elflink.c
+++ b/bfd/elflink.c
@@ -6608,7 +6608,6 @@ elf_gc_sweep_symbol (struct elf_link_hash_entry *h, void *data)
 
       inf = (struct elf_gc_sweep_symbol_info *) data;
       (*inf->hide_symbol) (inf->info, h, true);
-      h->def_regular = 0;
       h->ref_regular = 0;
       h->ref_regular_nonweak = 0;
     }
-- 
2.25.1

Re: [PATCH v1 1/1] bfd: elflink: upstream change to garbage collection sweep causes mb regression

[Michael Eager]

Hi Neal --

Thanks for the patch set.  I'll review them in the next couple days.

On 9/25/23 07:41, Neal Frager via Binutils wrote:
> Upstream change for PR13177 now clears the def_regular during gc_sweep of a
> section. (All other archs in binutils/bfd/elf32-*.c received an update
> to a warning about unresolvable relocations - this warning is not present
> in binutils/bfd/elf32-microblaze.c, but this warning check would not
> prevent the error being seen).
> 
> The visible issue with this change is when running a c++ application
> in Petalinux which links libstdc++.so for exception handling it segfaults
> on execution.
> 
> This does not occur if static linking libstdc++.a, so its during the
> relocations for a shared lib with garbage collection this occurs.
> 
> Signed-off-by: Neal Frager <neal.frager@amd.com>
> ---
>   bfd/elflink.c | 1 -
>   1 file changed, 1 deletion(-)
> 
> diff --git a/bfd/elflink.c b/bfd/elflink.c
> index ca162145f7e..0524019641e 100644
> --- a/bfd/elflink.c
> +++ b/bfd/elflink.c
> @@ -6608,7 +6608,6 @@ elf_gc_sweep_symbol (struct elf_link_hash_entry *h, void *data)
>   
>         inf = (struct elf_gc_sweep_symbol_info *) data;
>         (*inf->hide_symbol) (inf->info, h, true);
> -      h->def_regular = 0;
>         h->ref_regular = 0;
>         h->ref_regular_nonweak = 0;
>       }

Re: [PATCH v1 1/1] bfd: elflink: upstream change to garbage collection sweep causes mb regression

[Michael Eager]

On 9/25/23 07:41, Neal Frager via Binutils wrote:
> Upstream change for PR13177 now clears the def_regular during gc_sweep of a
> section. (All other archs in binutils/bfd/elf32-*.c received an update
> to a warning about unresolvable relocations - this warning is not present
> in binutils/bfd/elf32-microblaze.c, but this warning check would not
> prevent the error being seen).
> 
> The visible issue with this change is when running a c++ application
> in Petalinux which links libstdc++.so for exception handling it segfaults
> on execution.
> 
> This does not occur if static linking libstdc++.a, so its during the
> relocations for a shared lib with garbage collection this occurs.
> 
> Signed-off-by: Neal Frager <neal.frager@amd.com>
> ---
>   bfd/elflink.c | 1 -
>   1 file changed, 1 deletion(-)
> 
> diff --git a/bfd/elflink.c b/bfd/elflink.c
> index ca162145f7e..0524019641e 100644
> --- a/bfd/elflink.c
> +++ b/bfd/elflink.c
> @@ -6608,7 +6608,6 @@ elf_gc_sweep_symbol (struct elf_link_hash_entry *h, void *data)
>   
>         inf = (struct elf_gc_sweep_symbol_info *) data;
>         (*inf->hide_symbol) (inf->info, h, true);
> -      h->def_regular = 0;
>         h->ref_regular = 0;
>         h->ref_regular_nonweak = 0;
>       }

Neal --

It isn't clear what problem this patch addresses.

What update was applied to bfd/elf32-*.c for other archs?

This patch modifies code which is shared by all architectures, applied 
in a patch in 2011.  If there is a problem with this code, it seems that 
the failure would have been seen on other architectures long ago, not 
just MicroBlaze.

It seems likely that applying this patch would break other archs.

Please provide a test case which shows this problem.

-- 
Michael Eager