| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Some early revisions of the Cortex-A53 have an erratum (843419).
The details of the erratum are quite complex and involve dynamic
conditions. For the purposes of the workaround we have simplified
the static conditions to an ADRP in the last two instructions of a
4KByte page, followed within four instructions by a load/store dependent
on the ADRP.
This patch adds support to conservatively scan for and workaround the erratum.
There are two different workaround strategies used. The first is to rewrite
ADRP instructions which form part of an erratum sequence with an ADR instruction.
In situations where the ADR provides insufficient offset the dependent
load or store instruction from the sequence is moved to a stub section
and branches are inserted from the original sequence to the relocated
instruction and back again.
Stub section sizes are rounded up to a multiple of 4096 in order to ensure
that the act of inserting work around stubs does not create more errata sequences.
Workaround stubs are always inserted into the stub section associated
with the input section containing the erratum sequence. This ensures that the
fully relocated form of the veneered load store instruction is available at the
point in time when the stub section is written.
2015-03-30 Tejas Belagod <tejas.belagod@arm.com>
Marcus Shawcroft <marcus.shawcroft@arm.com>
Jiong Wang <jiong.wang@arm.com>
* bfd-in.h (bfd_elf64_aarch64_set_options)
(bfd_elf32_aarch64_set_options): Add parameter.
* bfd-in2.h: Regenerated.
* elfnn-aarch64.c (aarch64_erratum_843419_stub)
(_bfd_aarch64_adrp_p, _bfd_aarch64_erratum_843419_sequence_p)
(_bfd_aarch64_erratum_843419_stub_name)
(_bfd_aarch64_erratum_843419_fixup)
(_bfd_aarch64_erratum_843419_scan)
(_bfd_aarch64_erratum_843419_branch_to_stub)
(_bfd_aarch64_erratum_843419_p): Define.
(enum elf_aarch64_stub_type): Define
aarch64_stub_erratum_843419_veneer.
(struct elf_aarch64_stub_hash_entry): Define adrp_offset.
(struct elf_aarch64_link_hash_table): Define fix_erratum_843419
and fix_erratum_843419_adr.
(stub_hash_newfunc): Initialize adrp_offset;
(_bfd_aarch64_add_stub_entry_after): Define.
(aarch64_map_one_stub, aarch64_build_one_stub)
(aarch64_size_one_stub): Handle
aarch64_stub_erratum_843419_veneer.
(_bfd_aarch64_resize_stubs): Round stub section size.
(elfNN_aarch64_size_stubs): Add scan for 843419.
(bfd_elfNN_aarch64_set_options): Add parameter. Initialize
fix_erratum_843419 and fix_erratum_843419_adr.
(struct erratum_835769_branch_to_stub_data): Add info.
(elfNN_aarch64_write_section): Initialise info. Handle 843419.
(elfNN_aarch64_size_dynamic_sections): Handle 843419.
* elfxx-aarch64.c (_bfd_aarch64_decode_adrp_imm)
(_bfd_aarch64_sign_extend): Define.
(reencode_adr_imm): Remove static. Rename to:
(_bfd_aarch64_reencode_adr_imm): Define.
(_bfd_aarch64_elf_put_addend): Call _bfd_aarch64_reencode_adr_imm.
* elfxx-aarch64.h (AARCH64_ADR_OP, AARCH64_ADRP_OP)
(AARCH64_ADRP_OP_MASK, _bfd_aarch64_sign_extend)
(_bfd_aarch64_decode_adrp_imm, _bfd_aarch64_reencode_adr_imm):
Define.
ld/testsuite:
2015-03-24 Tejas Belagod <tejas.belagod@arm.com>
* emultempl/aarch64elf.em
(aarch64_elf_create_output_section_statements): Add parameter in
bfd_elf${ELFSIZE}_aarch64_set_options call.
(OPTION_FIX_ERRATUM_843419): Define.
(PARSE_AND_LIST_LONGOPTS): Add fix-cortex-a53-843419.
(PARSE_AND_LIST_ARGS_CASES): Add OPTION_FIX_ERRATUM_843419.
Change-Id: I758b04c90d431f4e72a978871d5a66676c413d92
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Ensure that injection of a stub section does not break a link where
there is an xpectation that flow of control can pass from one input
section to another simply by linking the input sections in series.
The solution here is to allow stub sections to be inserted after any
input section (existing behaviour), but inject an additional branch at
the start of each stub section such that control flow falling into the
stub section will branch over the stub section.
Change-Id: I8ccadcfb2f43e6409cb5a649087a47c0c8826b15
|
| |
|
|
| |
Change-Id: Idc292f60196ed4bf417eb77acc239761666adb8b
|
| |
|
|
| |
Change-Id: Ib6171d8646bc0e825a5038c7087acc534a8612a9
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Change the behaviour of section_group[] such that .stub_sec points to
the stub section attached to the indexed section rather than the stub
section attached to the link_section pointed to be the index section.
This provides a mechanism to get to the stub section following any
input section. While still allowing the section grouping mechanism to
find the section group stub section associated with an input section
by first following the link_sec pointer.
Change-Id: I19bce3feea77bbc71c04f2d8b3a204c527a21d9a
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This patch recognizes that we only need to perform one scan for the
835769 errata and that this scan can take place before we insert
branch stubs. The erratum scan code is relocated and adjusted to
create stub entries directly rather than populating an intermediate
representation. Since stub entries are created immediately we can
drop the adhoc stub size adjustment code and allow the generic stub
sizing code to deal with 835769 stub entries.
This patch restructures the code but does not change the workaround
used to deal with erratum 83679, the exact placement of workaround
stubs in the final image may change slightly after this patch due to
stubs being created in a different order.
Change-Id: I1ac3b2b7a8a5faf678a1e375f50f63ab6fd22d1f
|
| |
|
|
|
|
|
|
| |
Adjust the 835769 workaround code to use
_bfd_aarch64_add_stub_entry_in_group rather than inspect the
underlying section_group structure directly.
Change-Id: I26e7793af2a8bd6ab57c2b958cec12565d280a8f
|
| |
|
|
|
|
|
|
|
| |
bfd/ChangeLog
* elfnn-aarch64.c (_bfd_aarch64_create_or_find_stub_sec): Add
missing ';'.
Change-Id: I174cb391792c428521a9334eb166cdc8c979b12f
|
| | |
|
| |
|
|
| |
Change-Id: I3e92b1a8cf997ea7eb902bc38d5163cf919789e0
|
| |
|
|
| |
Change-Id: I6f01f9c9c3a64ed8ed52486f0f6f66f66b85a317
|
| |
|
|
|
|
| |
elf_aarch64_create_or_find_stub_sec and elfNN_aarch64_add_stub
Change-Id: I8e5ce0f4dd3aa322f97fa09d42c6f95e0d45118d
|
| |
|
|
| |
Change-Id: Ia79975e3ce9f1eca6ae092b4c9e2d92168fde1a0
|
| |
|
|
| |
Change-Id: I0454073cfad254ee5e28c5d597d615abdef9331d
|
| |
|
|
| |
Change-Id: I5e0be2e412d13fe4a5d738d7949bfb11909783d0
|
| |
|
|
| |
Change-Id: Iea5ae8c19790beebc6188b094895cbe7c7c221c5
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The BFD/LD patch for Cortex-A53 erratum 835769
(https://sourceware.org/ml/binutils/2014-10/msg00199.html)
does not handle a particular case of the AArch64 ELF ABI where mapping
symbols are allowed to be unordered in the symbol table (not in address order).
The unordering causes section maps to be traversed with incorrect span boundaries
(in the erratum scanning function) which causes memory faults. The attached
patch fixes this issue by ordering the section maps by their 'vma' before
starting to traverse them.
While this is not an issue with a the GNU toolchain, it is a potential issue
with Clang/LLVM. We have observed at least one case where LLVM generates an
ELF object with mapping symbols unordered in the symbol table and causes a fault.
We have been unable to construct a test case with the GNU toolchain. We have verified
by manual inspection the correctness of the traversal with this patch for an
LLVM-generated ELF object which triggered this issue. This patch has been bootstrapped
on aarch64-linux and regressed.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
local patches will be re-applied later
commit a30720e3e633f275250e26f85ccae5dbdddfb6c6
Author: Alan Modra <amodra@gmail.com>
Date: Wed Nov 19 10:30:16 2014 +1030
daily update
Change-Id: Ieb2a3f4dd2ecb289ac5305ff08d428b2847494ab
|
|
|
For MIPS -mmsa support
Change-Id: I08c4f002fa7b33dec85ed75956e6ab551bb03c96
|
