The problem is caused by the fact that gold is relocating the stubs
for an entire output section when it processes the relocations for a
particular input section that happened to be designated as the stub
table "owner". The Relocate_task for that input section may or may not
run before the Relocate_task for another input section that contains
the code that needs the erratum fix, but doesn't "own" the stub
table. If it runs before (or might even race with) that other task, it
ends up with a copy of the unrelocated original instruction.
In other words - when calling fix_errata() from
do_relocate_sections(), gold is going through the list of errata stubs
that are associated only with that object. This routine updates the
stored original instruction and replaces it in the output view with a
branch to the stub. Later, as gold is going through the object file's
input sections, it then checks for stub tables "owned" by each input
section, and writes out all the stubs from that stub table, regardless
of what object file each stub is associated with.
Fixed by relocating the erratum stub only after the corresponding
errata spot is fixed. That is to have fix_errata() call
Stub_table::relocate_erratum_stub() for each stub.
gold/ChangeLog
2017-07-06 Han Shen <shenhan@google.com>
PR gold/21491
* aarch64.cc (Erratum_stub::invalidate_erratum_stub): New method.
(Erratum_stub::is_invalidated_erratum_stub): New method.
(Stub_table::relocate_reloc_stub): Renamed from "relocate_stub".
(Stub_table::relocate_reloc_stubs): Renamed from "relocate_stubs".
(Stub_table::relocate_erratum_stub): New method.
(AArch64_relobj::fix_errata_and_relocate_erratum_stubs): Renamed from
"fix_errata".
(Target_aarch64::relocate_reloc_stub): Renamed from "relocate_stub".
This file contains invisible Unicode characters that are indistinguishable to humans but may be processed differently by a computer. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
gold is an ELF linker. It is intended to have complete support for
ELF and to run as fast as possible on modern systems. For normal use
it is a drop-in replacement for the older GNU linker.
gold is part of the GNU binutils. See ../binutils/README for more
general notes, including where to send bug reports.
gold was originally developed at Google, and was contributed to the
Free Software Foundation in March 2008. At Google it was designed by
Ian Lance Taylor, with major contributions by Cary Coutant, Craig
Silverstein, and Andrew Chatham.
The existing GNU linker manual is intended to be accurate
documentation for features which gold supports. gold supports most of
the features of the GNU linker for ELF targets. Notable
omissions--features of the GNU linker not currently supported in
gold--are:
* MRI compatible linker scripts
* cross-reference reports (--cref)
* various other minor options
Notes on the code
=================
These are some notes which may be helpful to people working on the
source code of gold itself.
gold is written in C++. It is a GNU program, and therefore follows
the GNU formatting standards as modified for C++. Source documents in
order of decreasing precedence:
http://www.gnu.org/prep/standards/
http://gcc.gnu.org/onlinedocs/libstdc++/manual/source_code_style.html
http://www.zembu.com/eng/procs/c++style.html
The linker is intended to have complete support for cross-compilation,
while still supporting the normal case of native linking as fast as
possible. In order to do this, many classes are actually templates
whose parameter is the ELF file class (e.g., 32 bits or 64 bits). The
C++ code is the same, but we don't pay the execution time cost of
always using 64-bit integers if the target is 32 bits. Many of these
class templates also have an endianness parameter: true for
big-endian, false for little-endian.
The linker is multi-threaded. The Task class represents a single unit
of work. Task objects are stored on a single Workqueue object. Tasks
communicate via Task_token objects. Task_token objects are only
manipulated while holding the master Workqueue lock. Relatively few
mutexes are used.
Build requirements
==================
The gold source code uses templates heavily. Building it requires a
recent version of g++. g++ 4.0.3 and 4.1.3 are known to work. g++
3.2, 3.4.3, and 4.1.2 are known to fail.
The linker script parser uses features which are only in newer
versions of bison. bison 2.3 is known to work. bison 1.26 is known
to fail. If you are building gold from an official binutils release,
the bison output should already be included.
Copyright (C) 2012-2017 Free Software Foundation, Inc.
Copying and distribution of this file, with or without modification,
are permitted in any medium without royalty provided the copyright
notice and this notice are preserved.
df2f63a6a0