Fix problem where erratum stubs are not always applied.
authorCary Coutant <ccoutant@gmail.com>
Thu, 30 Nov 2017 21:44:28 +0000 (13:44 -0800)
committerCary Coutant <ccoutant@gmail.com>
Thu, 30 Nov 2017 23:08:21 +0000 (15:08 -0800)
commite0feb133429709603eeeb382c1ebb6edd0a886aa
tree376441e0041c25f2cbf68b53778c64c4b1dc0ed1
parent036aae79303c43c56b604bc215ce4a70fc2a6474
Fix problem where erratum stubs are not always applied.

I checked over the results of applying --fix-cortex-a53-843419 to
a very large program (gitit) with two stub tables and thousands
of erratum fixes. I noticed that all the erratum_stubs were being
created but about 1/3 of them were being skipped over by
fix_errata_and_relocate_erratum_stubs(). By skipped over I mean
no branch relocation or adrp -> adr transformation was applied to
the erratum address, leaving the erratum_stub unreachable, and
with a branch with a 0 immediate.

The root cause of the skipped over erratum_stubs is
Erratum_stub::invalidate_erratum_stub() that is used to set
relobj_ to NULL when an erratum_stub has been processed.
Unfortunately relobj_ is used in operator<() so altering relobj
makes the results from erratum_stubs_.lower_bound() as used in
find_erratum_stubs_for_input_section() unreliable.

2017-11-30  Peter Smith  <peter.smith@linaro.org>
    Cary Coutant  <ccoutant@gmail.com>

gold/
PR gold/20765
* aarch64.cc (Erratum_stub::invalidate_erratum_stub): Use erratum_insn_
instead of relobj_ to invalidate the stub.
(Erratum_stub::is_invalidated_erratum_stub): Likewise.
gold/ChangeLog
gold/aarch64.cc