[PATCH] PR gold/20462: Fix bogus layout on ARM with linker script using PHDRS clause

[PATCH] PR gold/20462: Fix bogus layout on ARM with linker script using PHDRS clause

[Roland McGrath]

See https://sourceware.org/bugzilla/show_bug.cgi?id=20462 for the test
case.

I'm not at all sure this fix is correct.  It fixes both the minimal
test case I put in the bug, and my real case (which uses the same
linker script).  It causes no regressions in 'make check-gold' in
a trunk build with --target=arm-eabi.

Here's what I figured out:
* On the first pass, the layout comes out just fine.
* ARM always tries relaxation (unless -r).
* In Target_arm::do_relax it calls Target_arm::fix_exidx_coverage and
  decides that constituted some relaxation happening, so a second pass
  always happens when there are any .ARM.exidx sections.
* In the second pass, Script_sections::expected_segment_count sees
  this->segments_created_ and short-circuits, so it returns 0 instead
  of this->phdrs_elements_->size().
* Ergo, SIZEOF_HEADERS omits the phdrs in the second pass and
  addresses go backwards, making everything unhappy.

Script_sections::expected_segment_count is the only thing that tests
this->segments_created_.  It's not really clear to me if this ought
to be part of the state that Script_sections::release_segments (sole
caller Layout::clean_up_after_relaxation) resets, as my change here
does.  An alternate fix would be to reverse the order of the first two
checks in Script_sections::expected_segment_count so that it always
returns this->phdrs_elements_->size() when this->saw_phdrs_clause().
(I haven't tried that on my test cases.)

If this is indeed a good fix, may I commit it to trunk and 2.27?
If not, is my other suggestion right or is there a different fix?


Thanks,
Roland


gold/
2016-08-11  Roland McGrath  <roland@hack.frob.com>

        PR gold/20462
        * script-sections.cc (Script_sections::release_segments):
        Reset this->segments_created_.

diff --git a/gold/script-sections.cc b/gold/script-sections.cc
index 96c68de..bf25391 100644
--- a/gold/script-sections.cc
+++ b/gold/script-sections.cc
@@ -4459,6 +4459,7 @@ Script_sections::release_segments()
           ++p)
        (*p)->release_segment();
     }
+  this->segments_created_ = false;
 }

 // Print the SECTIONS clause to F for debugging.

Re: [PATCH] PR gold/20462: Fix bogus layout on ARM with linker script using PHDRS clause

[Cary Coutant]

> I'm not at all sure this fix is correct.  It fixes both the minimal
> test case I put in the bug, and my real case (which uses the same
> linker script).  It causes no regressions in 'make check-gold' in
> a trunk build with --target=arm-eabi.
>
> Here's what I figured out:
> * On the first pass, the layout comes out just fine.
> * ARM always tries relaxation (unless -r).
> * In Target_arm::do_relax it calls Target_arm::fix_exidx_coverage and
>   decides that constituted some relaxation happening, so a second pass
>   always happens when there are any .ARM.exidx sections.
> * In the second pass, Script_sections::expected_segment_count sees
>   this->segments_created_ and short-circuits, so it returns 0 instead
>   of this->phdrs_elements_->size().
> * Ergo, SIZEOF_HEADERS omits the phdrs in the second pass and
>   addresses go backwards, making everything unhappy.
>
> Script_sections::expected_segment_count is the only thing that tests
> this->segments_created_.  It's not really clear to me if this ought
> to be part of the state that Script_sections::release_segments (sole
> caller Layout::clean_up_after_relaxation) resets, as my change here
> does.  An alternate fix would be to reverse the order of the first two
> checks in Script_sections::expected_segment_count so that it always
> returns this->phdrs_elements_->size() when this->saw_phdrs_clause().
> (I haven't tried that on my test cases.)
>
> If this is indeed a good fix, may I commit it to trunk and 2.27?
> If not, is my other suggestion right or is there a different fix?

Yes, I think this is the right fix. OK for trunk and 2.27 branch.

Thanks!

-cary

Re: [PATCH] PR gold/20462: Fix bogus layout on ARM with linker script using PHDRS clause

[Roland McGrath]

On Fri, Aug 12, 2016 at 9:54 AM, Cary Coutant <ccoutant@gmail.com> wrote:
> Yes, I think this is the right fix. OK for trunk and 2.27 branch.

Committed to both.


Thanks,
Roland