In addition to the existing relocs we need two more to mark all
instructions in the call sequence, PLTCALL on the call itself (plus
the toc restore insn for ppc64), and PLTSEQ on others. All
relocations in a particular sequence have the same symbol.
Example ppc64 ELFv2 assembly:
.reloc .,R_PPC64_PLTSEQ,puts
std 2,24(1)
addis 12,2,puts@plt@ha # .reloc .,R_PPC64_PLT16_HA,puts
ld 12,puts@plt@l(12) # .reloc .,R_PPC64_PLT16_LO_DS,puts
.reloc .,R_PPC64_PLTSEQ,puts
mtctr 12
.reloc .,R_PPC64_PLTCALL,puts
bctrl
ld 2,24(1)
Example ppc32 -fPIC assembly:
addis 12,30,puts+32768@plt@ha # .reloc .,R_PPC_PLT16_HA,puts+0x8000
lwz 12,12,puts+32768@plt@l # .reloc .,R_PPC_PLT16_LO,puts+0x8000
.reloc .,R_PPC_PLTSEQ,puts+32768
mtctr 12
.reloc .,R_PPC_PLTCALL,puts+32768
bctrl
Marking sequences like this allows the linker to convert them to nops
and a direct call if the target symbol turns out to be local.
When the call is __tls_get_addr, each relocation shown above is paired
with an R_PPC*_TLSLD or R_PPC*_TLSGD reloc to additionally mark the
sequence for possible TLS optimization. The TLSLD or TLSGD relocs are
emitted first.
include/
* elf/ppc.h (R_PPC_PLTSEQ, R_PPC_PLTCALL): Define.
* elf/ppc64.h (R_PPC64_PLTSEQ, R_PPC64_PLTCALL): Define.
bfd/
* elf32-ppc.c (ppc_elf_howto_raw): Add PLTSEQ and PLTCALL howtos.
(is_plt_seq_reloc): New function.
(ppc_elf_check_relocs): Handle PLTSEQ and PLTCALL relocs.
(ppc_elf_tls_optimize): Handle inline plt call sequence.
(ppc_elf_relax_section): Handle PLTCALL reloc.
(ppc_elf_relocate_section): Nop out inline plt call sequence when
resolving locally.
* elf64-ppc.c (ppc64_elf_howto_raw): Add R_PPC64_PLTSEQ and
R_PPC64_PLTCALL entries. Comment R_PPC64_TOCSAVE.
(has_tls_get_addr_call): Correct comment.
(is_branch_reloc): Add PLTCALL.
(is_plt_seq_reloc): New function.
(ppc64_elf_check_relocs): Handle PLT16_LO_DS reloc. Set
has_tls_reloc for R_PPC64_TLSGD and R_PPC64_TLSLD. Create plt
entry for R_PPC64_PLTCALL.
(ppc64_elf_tls_optimize): Handle inline plt call sequence.
(ppc_type_of_stub): Handle PLTCALL reloc.
(toc_adjusting_stub_needed): Likewise.
(ppc64_elf_relocate_section): Set "can_plt_call" for PLTCALL
reloc insn. Nop out inline plt call sequence when resolving
locally. Handle __tls_get_addr inline plt call optimization.
elfcpp/
* powerpc.h (R_POWERPC_PLTSEQ, R_POWERPC_PLTCALL): Define.
gold/
* powerpc.cc (Target_powerpc::Track_tls::maybe_skip_tls_get_addr_call):
Handle inline plt sequence relocs.
(Stub_table::Plt_stub_key::Plt_stub_key): Likewise.
(Target_powerpc::Scan::reloc_needs_plt_for_ifunc): Likewise.
(Target_powerpc::Relocate::relocate): Likewise.
Necessary if gcc is to use PLT16 relocs to implement -mlongcall, and
there isn't a good technical reason why local symbols should be
excluded from PLT16 support. Non-ifunc local symbol PLT entries go in
a separate section to other PLT entries. In a fixed position
executable they won't need to be relocated, and in a PIE or shared
library I chose to not implement lazy relocation.
bfd/
* elf64-ppc.c (LOCAL_PLT_ENTRY_SIZE): Define.
(struct ppc_stub_hash_entry): Add symtype field.
(PLT_KEEP): Define.
(struct ppc_link_hash_table): Add pltlocal and relpltlocal.
(create_linkage_sections): Create pltlocal and relpltlocal.
(ppc64_elf_check_relocs): Allow PLT relocs on local symbols.
Set PLT_KEEP.
(ppc64_elf_adjust_dynamic_symbol): Keep PLT entries for inline calls.
(allocate_dynrelocs): Allocate pltlocal and relpltlocal.
(ppc64_elf_size_dynamic_sections): Size pltlocal and relpltlocal.
Keep PLT entries for inline calls against locals.
(ppc_build_one_stub): Use pltlocal as appropriate.
(ppc_size_one_stub): Likewise.
(ppc64_elf_size_stubs): Set symtype.
(build_global_entry_stubs_and_plt): Init pltlocal and write
relpltlocal for globals.
(write_plt_relocs_for_local_syms): Likewise for local syms.
(ppc64_elf_relocate_section): Support PLT for local syms.
* elf32-ppc.c (PLT_KEEP): Define.
(struct ppc_elf_link_hash_table): Add pltlocal and relpltlocal.
(ppc_elf_create_glink): Create pltlocal and relpltlocal.
(ppc_elf_check_relocs): Allow PLT relocs on local symbols.
Set PLT_KEEP. Adjust update_local_sym_info call.
(ppc_elf_adjust_dynamic_symbol): Keep PLT entries for inline calls.
(allocate_dynrelocs): Allocate pltlocal and relpltlocal.
(ppc_elf_size_dynamic_sections): Size pltlocal and relpltlocal.
(ppc_elf_relocate_section): Support PLT16 relocs for local syms.
(write_global_sym_plt): Init pltlocal and write relpltlocal.
(ppc_finish_symbols): Likewise for locals.
ld/
* emulparams/elf32ppc.sh (OTHER_RELRO_SECTIONS_2): Add .branch_lt.
(OTHER_GOT_RELOC_SECTIONS): Add .rela.branch_lt.
* testsuite/ld-powerpc/elfv2so.d: Update for symbol/stub reordering.
* testsuite/ld-powerpc/relbrlt.d: Likewise.
* testsuite/ld-powerpc/relbrlt.s: Likewise.
* testsuite/ld-powerpc/tlsso.r: Likewise.
* testsuite/ld-powerpc/tlstocso.r: Likewise.
gold/
* powerpc.cc (Target_powerpc::lplt_): New variable.
(Target_powerpc::lplt_section): Associated accessor.
(Target_powerpc::plt_off): Handle local non-ifunc symbols.
(Target_powerpc::make_lplt_section): New function.
(Target_powerpc::make_local_plt_entry): New function.
(Powerpc_relobj::do_relocate_sections): Write out lplt.
(Output_data_plt_powerpc::first_plt_entry_offset): Zero for lplt.
(Output_data_plt_powerpc::add_local_entry): New function.
(Output_data_plt_powerpc::do_write): Ignore lplt.
(Target_powerpc::make_iplt_section): Make lplt first.
(Target_powerpc::make_brlt_section): Make .branch_lt relro.
(Target_powerpc::Scan::local): Handle PLT16 relocs.
The PowerPC64 ELFv2 ABI and the PowerPC SysV ABI support a number of
relocations that can be used to create and access a PLT entry.
However, the relocs are not well defined. The PLT16 family of relocs
talk about "the section offset or address of the procedure linkage
table entry". It's plain that we do need a relative address when PIC
as otherwise we'd have dynamic text relocations, but "section offset"
doesn't specify which section. The most obvious one, ".plt", isn't
that useful because there is no readily available way of addressing
the start of the ".plt" section. Much more useful would be "the
GOT/TOC-pointer relative offset of the procedure linkage table entry",
and I suppose you could argue that is a "section offset" of sorts.
For PowerPC64 it is better to use the same TOC-pointer relative
addressing even when non-PIC, since ".plt" may be located outside the
range of a 32-bit address. However, for ppc32 we do want an absolute
address when non-PIC as a GOT pointer may not be set up. Also, for
ppc32 PIC we have a similar situation to R_PPC_PLTREL24 in that the
GOT pointer is set to a location in the .got2 section and we need to
specify the .got2 offset in the PLT16 reloc addend.
This patch supports PLT16 relocations using these semantics. This is
not an ABI change for ppc32 since the relocations were not previously
supported by GNU ld, but is for ppc64 where some of the PLT16 relocs
were supported. I'm not particularly concerned since the old ppc64
PLT16 reloc semantics made them almost completely useless.
bfd/
* elf32-ppc.c (ppc_elf_check_relocs): Handle PLT16 relocs.
(ppc_elf_relocate_section): Likewise.
* elf64-ppc.c (ppc64_elf_check_relocs): Handle PLT16_LO_DS.
(ppc64_elf_relocate_section): Likewise. Correct PLT16
resolution to plt entry relative to toc pointer.
gold/
* powerpc.cc (Target_powerpc::plt_off): New functions.
(is_plt16_reloc): New function.
(Stub_table::plt_off): Use Target_powerpc::plt_off.
(Stub_table::plt_call_size): Use plt_off.
(Stub_table::do_write): Likewise.
(Target_powerpc::Scan::get_reference_flags): Return RELATIVE_REF
for PLT16 relocations.
(Target_powerpc::Scan::reloc_needs_plt_for_ifunc): Return true
for PLT16 relocations.
(Target_powerpc::Scan::global): Make a PLT entry for PLT16 relocations.
(Target_powerpc::Relocate::relocate): Support PLT16 relocations.
(Powerpc_scan_relocatable_reloc::global_strategy): Return RELOC_SPECIAL
for ppc32 plt16 relocs.
Relocations referring to discarded sections are now treated as errors
instead of warnings.
Also with this patch, we will now print the section group signature and the
object file with the prevailing definition of that group along with the
name of the symbol that the relocation is referring to. This additional
information should be much more useful to anyone trying to track down
the source of such errors.
To do so, we now map each discarded section to the Kept_section info in
the Layout class, and defer the logic that maps a discarded section to
its counterpart in the kept group. This gives us the information we need
to identify the signature symbol given the discarded section, and the
name of the object file that provided the prevailing (i.e., first)
definition of that group.
gold/
* object.cc (Sized_relobj_file::include_section_group): Store
reference to Kept_section info for discarded comdat sections
regardless of size. Move size checking to map_to_kept_section.
(Sized_relobj_file::include_linkonce_section): Likewise.
(Sized_relobj_file::map_to_kept_section): Add section name parameter.
Insert size checking logic from above functions.
(Sized_relobj_file::find_kept_section_object): New method.
(Sized_relobj_file::get_symbol_name): New method.
* object.h (Sized_relobj_file::map_to_kept_section): Add section_name
parameter. Adjust all callers.
(Sized_relobj_file::find_kept_section_object): New method.
(Sized_relobj_file::get_symbol_name): New method.
(Sized_relobj_file::Kept_comdat_section): Replace object and shndx
fields with sh_size, kept_section, symndx, and is_comdat fields.
(Sized_relobj_file::set_kept_comdat_section): Replace kept_object
and kept_shndx parameters with is_comdat, symndx, sh_size, and
kept_section.
(Sized_relobj_file::get_kept_comdat_section): Likewise.
* target-reloc.h (enum Comdat_behavior): Change CB_WARNING to CB_ERROR.
Adjust all references.
(issue_undefined_symbol_error): New function template.
(relocate_section): Pass section name to map_to_kept_section.
Move discarded section code to new function above.
* aarch64.cc (Target_aarch64::scan_reloc_section_for_stubs): Move
declaration for gsym out one level. Call issue_discarded_error.
* arm.cc (Target_arm::scan_reloc_section_for_stubs): Likewise.
* powerpc.cc (Relocate_comdat_behavior): Change CB_WARNING to CB_ERROR.
During a -r (or --emit-relocs) link, if two sections had the same name but
different section types, gold would put relocations for both sections into
the same relocation section even though the data sections remained separate.
For .eh_frame sections, when one section is PROGBITS and another is
X86_64_UNWIND, we really should be using the UNWIND section type and
combining the sections anyway. For other sections, we should be
creating one relocation section for each output data section.
gold/
PR gold/23016
* incremental.cc (can_incremental_update): Check for unwind section
type.
* layout.h (Layout::layout): Add sh_type parameter.
* layout.cc (Layout::layout): Likewise.
(Layout::layout_reloc): Create new output reloc section if data
section does not already have one.
(Layout::layout_eh_frame): Check for unwind section type.
(Layout::make_eh_frame_section): Use unwind section type for .eh_frame
and .eh_frame_hdr.
* object.h (Sized_relobj_file::Shdr_write): New typedef.
(Sized_relobj_file::layout_section): Add sh_type parameter.
(Sized_relobj_file::Deferred_layout::Deferred_layout): Add sh_type
parameter.
* object.cc (Sized_relobj_file::check_eh_frame_flags): Check for
unwind section type.
(Sized_relobj_file::layout_section): Add sh_type parameter; pass it
to Layout::layout.
(Sized_relobj_file::do_layout): Make local copy of sh_type.
Force .eh_frame sections to unwind section type.
Pass sh_type to layout_section.
(Sized_relobj_file<size, big_endian>::do_layout_deferred_sections):
Pass sh_type to layout_section.
* output.cc (Output_section::Output_section): Initialize reloc_section_.
* output.h (Output_section::reloc_section): New method.
(Output_section::set_reloc_section): New method.
(Output_section::reloc_section_): New data member.
* target.h (Target::unwind_section_type): New method.
(Target::Target_info::unwind_section_type): New data member.
* aarch64.cc (aarch64_info): Add unwind_section_type.
* arm.cc (arm_info, arm_nacl_info): Likewise.
* i386.cc (i386_info, i386_nacl_info, iamcu_info): Likewise.
* mips.cc (mips_info, mips_nacl_info): Likewise.
* powerpc.cc (powerpc_info): Likewise.
* s390.cc (s390_info): Likewise.
* sparc.cc (sparc_info): Likewise.
* tilegx.cc (tilegx_info): Likewise.
* x86_64.cc (x86_64_info, x86_64_nacl_info): Likewise.
* testsuite/Makefile.am (pr23016_1, pr23016_2): New test cases.
* testsuite/Makefile.in: Regenerate.
* testsuite/testfile.cc: Add unwind_section_type.
* testsuite/pr23016_1.sh: New test script.
* testsuite/pr23016_1a.s: New source file.
* testsuite/pr23016_1b.s: New source file.
* testsuite/pr23016_2.sh: New test script.
* testsuite/pr23016_2a.s: New source file.
* testsuite/pr23016_2b.s: New source file.
This reverts most of commit 1be5d8d3bb.
Left in place are addition of --no-plt-align to some ppc32 ld tests
and the ld.texinfo --no-plt-thread-safe fix.
Asking for ppc32 plt call stubs to be aligned at 32 byte boundaries
didn't quite work. For ld.bfd they were spaced 32 bytes apart, but
only started on a 16 byte boundary. ld.gold also didn't get it right.
Finding that bug made me check over the ppc64 plt stub alignment,
where I found that negative values for alignment (meaning align to
minimize boundary crossing) were not accepted. Since no one has
complained about that, I guess I could have removed the feature from
ld.bfd documentation, but I've opted instead to correct the code.
I've also added an optional alignment paramenter for ppc32
--plt-align, for some consistency with gold and ppc64 ld.bfd.
bfd/
* elf32-ppc.c (ppc_elf_create_glink): Correct alignment of .glink.
* elf64-ppc.c (ppc64_elf_size_stubs): Handle negative plt_stub_align.
(ppc64_elf_build_stubs): Likewise.
gold/
* powerpc.cc (param_plt_align): New function supplying default
--plt-align values. Use it..
(Stub_table::plt_call_align): ..here, and..
(Output_data_glink::global_entry_align): ..here.
(Stub_table::stub_align): Correct 32-bit minimum alignment.
ld/
* emultempl/ppc32elf.em: Support optional --plt-align arg.
* emultempl/ppc64elf.em: Support negative --plt-align arg.
This is in preparation for the next patch adding Spectre variant 2
mitigation for PowerPC and PowerPC64. Besides tidying code involved
in stub output (to reduce the number of places where bctr is output),
the patch adds some user visible features:
1) PowerPC64 ELFv2 global entry stubs now are aligned under the
control of --plt-align, with a default alignment of 32 bytes.
2) PowerPC64 __glink_PLTresolve is no longer padded out with nops.
3) PowerPC32 PLT stubs are aligned under the control of --plt-align,
with the default alignment being 16 bytes as before.
4) The PowerPC32 branch/nop table emitted before __glink_PLTresolve
is now smaller in many cases. It was sized incorrectly when the
__tls_get_addr_opt stub was used, and unnecessarily included space
for local ifuncs.
bfd/
* elf32-ppc.c (GLINK_ENTRY_SIZE): Add parameters, handle
__tls_get_addr_opt, and alignment sizing.
(TLS_GET_ADDR_GLINK_SIZE): Delete.
(is_nonpic_glink_stub): Don't use GLINK_ENTRY_SIZE.
(ppc_elf_get_synthetic_symtab): Recognize stubs spaced at 4, 6,
or 8 insns.
(ppc_elf_link_hash_table_create): Init new ppc_elf_params field.
(allocate_dynrelocs): Use new GLINK_ENTRY_SIZE.
(ppc_elf_size_dynamic_sections): Likewise. Size branch table
by PLT reloc count.
(write_glink_stub): Handle __tls_get_addr_opt stub.
Pad out to size given by GLINK_ENTRY_SIZE.
(ppc_elf_relocate_section): Adjust write_glink_stub call.
(ppc_elf_finish_dynamic_symbol): Likewise.
(ppc_elf_finish_dynamic_sections): Write PLTresolve without using
insn array since so many need rewriting.
* elf32-ppc.h (struct ppc_elf_params): Add plt_stub_align.
* elf64-ppc.c (GLINK_PLTRESOLVE_SIZE): Rename from
GLINK_CALL_STUB_SIZE. Add htab param and evaluate to size without
nops. Adjust all uses.
(ppc64_elf_get_synthetic_symtab): Don't use GLINK_CALL_STUB_SIZE
in glink_vma calculation.
(struct ppc_link_hash_table): Add global_entry section pointer.
(create_linkage_sections): Create separate section for global
entry stubs.
(PPC_LO, PPC_HI, PPC_HA): Move earlier.
(size_global_entry_stubs): Handle sizing for aligned stubs.
(ppc64_elf_size_dynamic_sections): Handle global_entry alloc,
and don't stash end of glink branch table in rawsize.
(ppc_build_one_stub): Rewrite stub size calculations.
(build_global_entry_stubs): Use new section.
(ppc64_elf_build_stubs): Don't pad __glink_PLTresolve with nops.
Build lazy link stubs out to end of section. Build global entry
stubs in new section.
gold/
* options.h (plt_align): Support for PowerPC32 too.
* powerpc.cc (Stub_table::stub_align): Heed --plt-align for 32-bit.
(Stub_table::plt_call_size, branch_stub_size): Tidy.
(Stub_table::plt_call_align): Implement using stub_align.
(Output_data_glink::global_entry_align): New function.
(Output_data_glink::global_entry_off): New function.
(Output_data_glink::global_entry_address): Use global_entry_off.
(Output_data_glink::pltresolve_size): New function, replacing
pltresolve_size_ constant. Update all uses.
(Output_data_glink::add_global_entry): Align offset.
(Output_data_glink::set_final_data_size): Use global_entry_align.
(Stub_table::do_write): Don't pad __glink_PLTrelsolve with nops.
Tidy stub output. Use global_entry_off.
ld/
* emultempl/ppc32elf.em (params): Init new field.
(enum ppc32_opt): New enum to define OPTION_* values. Add
OPTION_PLT_ALIGN and OPTION_NO_PLT_ALIGN.
(PARSE_AND_LIST_LONGOPTS): Handle new options.
(PARSE_AND_LIST_ARGS_CASES): Likewise.
(PARSE_AND_LIST_OPTIONS): Likewise. Break up help output.
* emultempl/ppc64elf.em (ppc_add_stub_section): Init alignment
correctly for negative --plt-stub-align.
* testsuite/ld-powerpc/elfv2exe.d,
* testsuite/ld-powerpc/elfv2so.d,
* testsuite/ld-powerpc/relbrlt.d,
* testsuite/ld-powerpc/relbrlt.s,
* testsuite/ld-powerpc/tlsexe.d,
* testsuite/ld-powerpc/tlsexe.r,
* testsuite/ld-powerpc/tlsexe32.d,
* testsuite/ld-powerpc/tlsexe32.g,
* testsuite/ld-powerpc/tlsexe32.r,
* testsuite/ld-powerpc/tlsexetoc.d,
* testsuite/ld-powerpc/tlsexetoc.r,
* testsuite/ld-powerpc/tlsopt5_32.d,
* testsuite/ld-powerpc/tlsso.d,
* testsuite/ld-powerpc/tlstocso.d: Update for changed stub order.
We never need to resolve_forwards() a symbol found by hash table lookup
such as target->tls_get_addr_opt() but we do potentially need to do so
for random symbols seen on relocs. So these calls were in the wrong
order, resulting in missing stubs and an assertion failure.
PR 22602
* powerpc.cc (Target_powerpc::Branch_info::mark_pltcall): Resolve
forwards before replacing __tls_get_addr.
(Target_powerpc::Branch_info::make_stub): Likewise.
The fix for PR 19291 broke some other cases where -r is used with scripts,
as reported in PR 22266. The original fix for PR 22266 ended up breaking
many cases for REL targets, where the addends are stored in the section data,
and are not being adjusted properly.
The problem was basically that in a relocatable output file (ET_REL),
symbol values are supposed to be relative to the start address of their
section. Usually in a relocatable file, all sections start at 0, so the
failure to get this right is often irrelevant, but with a linker script,
we occasionally see an output section whose starting address is not 0,
and gold would occasionally write a symbol with its relocated value instead
of its section-relative value.
This patch reverts the recent fix for PR 22266 as well as my original fix
for PR 19291. The original fix moved the symbol value adjustment to
write_local_symbols, but neglected to undo a few places where the adjustment
was also being applied, resulting in an occasional double adjustment. The
more recent fix removed those other adjustments, but then failed to
re-account for the adjustment when rewriting the relocations on REL targets.
With the old attempts reverted, we now apply the symbol value adjustment to
the one case that had been missed (non-section symbols in merge sections).
But now we also need to account for the adjustment when rewriting the addends
for RELA relocations.
gold/
PR gold/19291
PR gold/22266
* object.cc (Sized_relobj_file::compute_final_local_value_internal):
Revert changes from 2017-11-08 patch. Adjust symbol value in
relocatable links for non-section symbols.
(Sized_relobj_file::compute_final_local_value): Revert changes from
2017-11-08 patch.
(Sized_relobj_file::do_finalize_local_symbols): Likewise.
(Sized_relobj_file::write_local_symbols): Revert changes from
2015-11-25 patch.
* object.h (Sized_relobj_file::compute_final_local_value_internal):
Revert changes from 2017-11-08 patch.
* powerpc.cc (Target_powerpc::relocate_relocs): Adjust addend for
relocatable links.
* target-reloc.h (relocate_relocs): Adjust addend for relocatable links.
* testsuite/pr22266_a.c (hello): New function.
* testsuite/pr22266_main.c (main): Add test for merge sections.
* testsuite/pr22266_script.t: Add rule for .rodata.
Fixes a thinko. Given code that puts variables into the TOC (a bad
idea, but some see the TOC as a small data section) this bug could
result in an attempt to optimize a sequence that should not be
optimized.
* powerpc.cc (Target_powerpc::Scan::local): Correct dst_off
calculation for TOC16 relocs.
(Target_powerpc::Scan::global): Likewise.
gcc doesn't emit stack notes for ELFv1, since ELFv1 never needs an
executable stack. Note that ELFv1 is usually big-endian and ELFv2
little-endian, but the ABI is really orthogonal to endiannes.
* powerpc.cc (Target_powerpc<64,*>::powerpc_info): Set
is_default_stack_executable false.
ppc32, like many targets, defines the address of a function as the PLT
call stub code for functions referenced but not defined in a non-PIC
executable. ppc32 gold, unlike other targets, inherits the ppc64
multiple stub capability for dealing with very large binaries where
one set of stubs can't be reached from all code locations. This means
there can be multiple choices of address for a function, which might
cause function pointer comparison failures. So for ppc32, make
non-branch references always use the first stub group.
(PowerPC64 ELFv1 is always PIC so doesn't need to define the address
of an external function as the PLT stub. PowerPC64 ELFv2 needs a
special set of global entry stubs to serve as the address of external
functions, so it too is not affected by this bug.)
* powerpc.cc (Target_powerpc::Branch_info::make_stub): Put
stubs for ppc32 non-branch relocs in first stub table.
(Target_powerpc::Relocate::relocate): Resolve similarly.
In the TLS GD/LD to LE optimization, ld replaces a sequence like
addi 3,2,x@got@tlsgd R_PPC64_GOT_TLSGD16 x
bl __tls_get_addr(x@tlsgd) R_PPC64_TLSGD x
R_PPC64_REL24 __tls_get_addr
nop
with
addis 3,13,x@tprel@ha R_PPC64_TPREL16_HA x
addi 3,3,x@tprel@l R_PPC64_TPREL16_LO x
nop
When the tprel offset is small, this can be further optimized to
nop
addi 3,13,x@tprel
nop
bfd/
* elf64-ppc.c (struct ppc_link_hash_table): Add do_tls_opt.
(ppc64_elf_tls_optimize): Set it.
(ppc64_elf_relocate_section): Nop addis on TPREL16_HA, and convert
insn on TPREL16_LO and TPREL16_LO_DS relocs to use r13 when
addis would add zero.
* elf32-ppc.c (struct ppc_elf_link_hash_table): Add do_tls_opt.
(ppc_elf_tls_optimize): Set it.
(ppc_elf_relocate_section): Nop addis on TPREL16_HA, and convert
insn on TPREL16_LO relocs to use r2 when addis would add zero.
gold/
* powerpc.cc (Target_powerpc::Relocate::relocate): Nop addis on
TPREL16_HA, and convert insn on TPREL16_LO and TPREL16_LO_DS
relocs to use r2/r13 when addis would add zero.
ld/
* testsuite/ld-powerpc/tls.s: Add calls with tls markers.
* testsuite/ld-powerpc/tls32.s: Likewise.
* testsuite/ld-powerpc/powerpc.exp: Run tls marker tests.
* testsuite/ld-powerpc/tls.d: Adjust for TPREL16_HA/LO optimization.
* testsuite/ld-powerpc/tlsexe.d: Likewise.
* testsuite/ld-powerpc/tlsexetoc.d: Likewise.
* testsuite/ld-powerpc/tlsld.d: Likewise.
* testsuite/ld-powerpc/tlsmark.d: Likewise.
* testsuite/ld-powerpc/tlsopt4.d: Likewise.
* testsuite/ld-powerpc/tlstoc.d: Likewise.
This implements the special __tls_get_addr_opt call stub for powerpc
gold that returns __thread variable addresses without actually making
a call to __tls_get_addr in most cases. Shared libraries that are
loaded at program load time (ie. dlopen is not used) have a known
layout for their __thread variables, and thus DTPMOD64/DPTREL64 pairs
describing those variables can be set up by ld.so for the
__tls_get_addr_opt call stub fast exit.
Ref https://sourceware.org/ml/libc-alpha/2015-03/msg00626.html
I really, really wish I'd used a differently versioned __tls_get_addr
symbol than the base symbol to indicate glibc support for the
optimized call, rather than having glibc export __tls_get_addr_opt. A
lot of the messing around here, flipping symbols from __tls_get_addr
to __tls_get_addr_opt, is caused by that decision. About the only
benefit is that a user can see at a glance that their disassembled
code is calling __tls_get_addr via the fancy call stub.. Anyway, we
need references to __tls_get_addr to seem like they were to
__tls_get_addr_opt, and in cases like the tsan interceptor, a
definition of __tls_get_addr to seem like one of __tls_get_addr_opt
as well. That's the reason for Symbol::clear_in_reg and
Symbol_table::clone, and why symbols are substituted in Scan::global
and other places dealing with dynamic linking.
elfcpp/
* elfcpp.h (DT_PPC_OPT): Define.
* powerpc.h (PPC_OPT_TLS): Define.
gold/
* options.h (tls_get_addr_optimize): New option.
* symtab.h (Symbol::clear_in_reg, clone): New functions.
(Sized_symbol::clone): New function.
(Symbol_table::clone): New function.
* resolve.cc (Symbol::clone, Sized_symbol::clone): New functions.
* powerpc.cc (Target_powerpc::has_tls_get_addr_opt_,
tls_get_addr_, tls_get_addr_opt_): New vars.
(Target_powerpc::tls_get_addr_opt, tls_get_addr,
is_tls_get_addr_opt, replace_tls_get_addr,
set_has_tls_get_addr_opt, stk_linker): New functions.
(Target_powerpc::Track_tls::maybe_skip_tls_get_addr_call): Add
target param. Update callers. Compare symbols rather than names.
(Target_powerpc::do_define_standard_symbols): Init tls_get_addr_
and tls_get_addr_opt_.
(Target_powerpc::Branch_info::mark_pltcall): Translate tls_get_addr
sym to tls_get_addr_opt.
(Target_powerpc::Branch_info::make_stub): Likewise.
(Stub_table::define_stub_syms): Likewise.
(Target_powerpc::Scan::global): Likewise.
(Target_powerpc::Relocate::relocate): Likewise.
(add_3_12_2, add_3_12_13, bctrl, beqlr, cmpdi_11_0, cmpwi_11_0,
ld_11_1, ld_11_3, ld_12_3, lwz_11_3, lwz_12_3, mr_0_3, mr_3_0,
mtlr_11, std_11_1): New constants.
(Stub_table::eh_frame_added_): Delete.
(Stub_table::tls_get_addr_opt_bctrl_, plt_fde_len_, plt_fde_): New vars.
(Stub_table::init_plt_fde): New functions.
(Stub_table::add_eh_frame, replace_eh_frame): Move definition out
of line. Init and use plt_fde_.
(Stub_table::plt_call_size): Return size for tls_get_addr stub.
Extract alignment code to..
(Stub_table::plt_call_align): ..this new function. Adjust all callers.
(Stub_table::add_plt_call_entry): Set has_tls_get_addr_opt and
tls_get_addr_opt_bctrl, and align after that.
(Stub_table::do_write): Write out tls_get_addr stub.
(Target_powerpc::do_finalize_sections): Emit DT_PPC_OPT
PPC_OPT_TLS/PPC64_OPT_TLS bit.
(Target_powerpc::Relocate::relocate): Don't check for or modify
nop following bl for tls_get_addr stub.
This patch provides a flag for PowerPC64 ELFv2 use in class Symbol,
and modifies Sized_target::resolve to return whether the symbol has
been resolved. If not, normal processing continues. I use this for
PowerPC64 ELFv2 to keep track of whether a symbol has any definition
with non-zero localentry, in order to disable --plt-localentry for
that symbol.
PR 21847
* powerpc.cc (Target_powerpc::is_elfv2_localentry0): Test
non_zero_localentry.
(Target_powerpc::resolve): New function.
(powerpc_info): Set has_resolve for 64-bit.
* target.h (Sized_target::resolve): Return bool.
* resolve.cc (Symbol_table::resolve): Continue with normal
processing when target resolve returns false.
* symtab.h (Symbol::non_zero_localentry, set_non_zero_localentry):
New accessors.
(Symbol::non_zero_localentry_): New flag bit.
* symtab.cc (Symbol::init_fields): Init non_zero_localentry_.
There is a very small but non-zero probability that a stub group
contains stubs on one relax pass, but does not on the next. In that
case we would get an FDE covering a zero length address range.
(Actually, it's even worse. Alignment padding for stubs can mean the
address for the non-existent stubs is past the end of the original
section to which stubs are attached, and due to the way
do_plt_fde_location calculates the length we can get a negative
length.) Fixing this properly requires removing the FDE.
Also, I have been implementing the __tls_get_addr_opt support for
gold, and that stub needs something other than the default FDE. The
necessary FDE will depend on the offset to the __tls_get_addr_opt
stub, which of course can change during relaxation. That means at the
very least, rewriting the FDE on each pass, possibly changing the FDE
size. I think that is better done by completely recreating PLT
eh_frame FDEs.
* ehframe.cc (Fde::operator==): New.
(Cie::remove_fde, Eh_frame::remove_ehframe_for_plt): New.
* ehframe.h (Fde::operator==): Declare.
(Cie::remove_fde, Eh_frame::remove_ehframe_for_plt): Likewise.
* layout.cc (Layout::remove_eh_frame_for_plt): New.
* layout.h (Layout::remove_eh_frame_for_plt): Declare.
* powerpc.cc (Target_powerpc::do_relax): Remove old eh_frame FDEs.
(Stub_table::add_eh_frame): Delete eh_frame_added_ condition.
Don't add eh_frame for empty stub section.
(Stub_table::remove_eh_frame): New.
This adds a --no-tls-optimize option for people who want to keep
__tls_get_addr calls in an executable rather than optimizing such code
sequences to IE/LE.
Also tidy some formatting errors, rename a variable to better reflect
its use, and tweak two functions that create pairs of GOT entries to
first check whether the GOT entry already exists before potentially
inserting the header via reserve(2). Without the check it is possible
to waste one GOT entry.
* options.h (no_tls_optimize): New powerpc option.
* powerpc.cc (Target_powerpc::abiversion, set_abiversion): Formatting.
(Target_powerpc::stk_toc): Formatting, fix comment.
(Target_powerpc::Track_tls::tls_get_addr_state): Rename from
tls_get_addr.
(Target_powerpc::optimize_tls_gd, optimize_tls_ld, optimize_tls_ie):
Return TLSOPT_NONE when !tls_optimize.
(Target_powerpc::add_global_pair_with_rel): Check
for existing reloc before reserving.
(Target_powerpc::add_local_tls_pair): Likewise.
This makes ld warn about --plt-localentry if a version of glibc
without the necessary ld.so checks is detected, and revises the
documentation.
bfd/
* elf64-ppc.c (ppc64_elf_tls_setup): Warn on --plt-localentry
without ld.so checks.
gold/
* powerpc.cc (Target_powerpc::scan_relocs): Warn on --plt-localentry
without ld.so checks.
ld/
* ld.texinfo (plt-localentry): Revise.
The big comment in ppc64_elf_tls_setup says why. I've also added some
code to the bfd linker that catches the -lpthread -lc symbol
differences and disable generation of optimized call stubs even when
--plt-localentry is activated. Gold doesn't yet have that.
PR 21847
bfd/
* elf64-ppc.c (struct ppc_link_hash_entry): Add non_zero_localentry.
(ppc64_elf_merge_symbol): Set non_zero_localentry.
(is_elfv2_localentry0): Test non_zero_localentry.
(ppc64_elf_tls_setup): Default to --no-plt-localentry.
gold/
* powerpc.cc (Target_powerpc::scan_relocs): Default to
--no-plt-localentry.
ld/
* ld.texinfo (plt-localentry): Document.
My PPC64_OPT_LOCALENTRY patch of June 1, git commit f378ab099d, and
the later gold change, git commit 7ee7ff7015, added an insn in
__glink_PLTresolve which needs a corresponding adjustment in the
eh_frame info for asynchronous exceptions to unwind correctly.
It would have been OK for both ABIs to use +5 for the advance before
restore of LR, since we can put the DW_CFA_restore_extended on any
insn after the actual restore and before the r12/r0 copy is clobbered,
but it's slightly better to delay as much as possible. There are
then more addresses where fewer CFA program insns are executed.
bfd/
* elf64-ppc.c (ppc64_elf_size_stubs): Correct advance to
restore of LR.
gold/
* powerpc.cc (glink_eh_frame_fde_64v2): Correct advance to
restore of LR.
(glink_eh_frame_fde_64v1): Advance to restore of LR at latest
possible insn.
elfcpp/
* elfcpp.h (DT_PPC64_OPT): Define.
* powerpc.h (PPC64_OPT_TLS, PPC64_OPT_MULTI_TOC,
PPC64_OPT_LOCALENTRY): Define.
gold/
* options.h (General_options): Add plt_localentry.
* powerpc.cc (Target_powerpc::st_other): New function.
(Target_powerpc::plt_localentry0_, plt_localentry0_init_,
has_localentry0_): New vars.
(Target_powerpc::plt_localentry0, set_has_localentry0,
is_elfv2_localentry0): New functions.
(Target_powerpc::Branch_info::mark_pltcall): Don't set tocsave or
return true for localentry:0 calls.
(Stub_table::Plt_stub_ent::localentry0_): New var.
(Stub_table::add_plt_call_entry): Set localentry0_ and has_localentry0_.
Don't set r2save_ for localentry:0 calls.
(Output_data_glink::do_write): Save r2 in __glink_PLTresolve for elfv2.
(Target_powerpc::scan_relocs): Default plt_localentry0_.
(Target_powerpc::do_finalize_sections): Set DT_PPC64_OPT.
(Target_powerpc::Relocate::relocate): Don't require nop following
calls for localentry:0 plt calls, and don't change nop.
This adds support to gold for the tocsave relocs already supported by
ld.bfd. R_PPC64_TOCSAVE relocs are part of a scheme to move r2 saves
to the prologue of a function rather than in each plt call stub. We
don't want a compiler to always emit the r2 save, as this would be
wasted if the calls turned out to be local. See the tocsave*.s in
ld/testsuite/ld-powerpc/.
* powerpc.cc (Target_powerpc::tocsave_loc_): New var.
(Target_powerpc::mark_pltcall, add_tocsave, tocsave_loc): New functions.
(Target_powerpc::Branch_info::tocsave_): New var.
(Target_powerpc::Branch_info::mark_pltcall): New function.
(Target_powerpc::Branch_info::make_stub): Pass tocsave_ to
add_plt_call_entry.
(Stub_table::Plt_stub_ent): Make public. Add r2save_.
(Stub_table::add_plt_call_entry): Add bool tocsave_ param. Set
r2save_.
(Stub_table::find_plt_call_entry): Return Plt_stub_ent*. Adjust
use throughout.
(Stub_table::do_write): Conditionally output r2 save in plt stubs.
(Target_powerpc::Scan::local): Handle R_PPC64_TOCSAVE.
(Target_powerpc::Scan::global): Likewise.
(Target_powerpc::Relocate::relocate): Skip r2 save in plt call stub
with tocsave reloc. Replace header tocsave nop with r2 save.
* symtab.h (struct Symbol_location_hash): Make public.
I was lazy when adding indx_ to Plt_stub_ent. The field isn't part of
the key, so ought to be part of the mapped type. Make it so.
* powerpc.cc (Plt_stub_key): Rename from Plt_stub_ent. Remove indx_.
(Plt_stub_key_hash): Rename from Plt_stub_ent_hash.
(struct Plt_stub_ent): New.
(Plt_stub_entries): Map from Plt_stub_key to Plt_stub_ent. Adjust
use throughout file.
If two objects are compiled with -fPIC or -fPIE and call the same
function, two different PLT entries are created, one for each object,
but the same stub symbol name is used for both.
* powerpc.cc (Stub_table::define_stub_syms): Always include object's
uniq_ value.
Doesn't yet trim off the unused TOC entries.
* powerpc.cc (class Powerpc_copy_relocs): New.
(Powerpc_copy_relocs::emit): New function.
(Powerpc_relobj::relatoc_, toc_, no_toc_opt_): New variables.
(Powerpc_relobj::toc_shndx, set_no_toc_opt, no_toc_opt): New inlines.
(Powerpc_relobj::do_relocate_sections): New function.
(Powerpc_relobj::make_toc_relative): Likewise.
(Powerpc_relobj::do_find_special_sections): Stash away .rela.toc
and .toc too.
(ok_lo_toc_insn): Move earlier, and handle more insns.
(Target_powerpc::Scan::local): If optimizing toc accesses, set
no_toc_opt for entries we can't edit. Check insn validity.
Emit "toc optimization is not supported" warning, downgraded
from error.
(Target_powerpc::Scan::global): Likewise.
(Target_powerpc::Relocate::relocate): Edit TOC indirect code
to TOC relative. Don't emit "toc optimization is not supported"
error here.
Added just to accept, and ignore. gcc since 2015-10-21, when
configured with --enable-secureplt passes this option to the linker.
As powerpc gold cannot link --bss-plt code successfully, gold needs to
accept the option or the gcc specs file needs to be changed.
The patch also make gold detect --bss-plt code and error out rather
than producing a binary that crashes.
* options.h: Add --secure-plt option.
* powerpc.cc (Target_powerpc::Scan::local): Detect and error
on -fPIC -mbss-plt code.
(Target_powerpc::Scan::global): Likewise.
Plus some paranoia in symval_for_branch. We shouldn't get there with
dynamic symbols, but if we ever did the static_cast to Powerpc_relobj
would be wrong.
* powerpc.cc: Use shorter equivalent elfcpp typedef for
Reltype and reloc_size throughout.
(Target_powerpc::symval_for_branch): Exclude dynamic symbols.
(Target_powerpc::Scan::local): Use local var r_sym.
(Target_powerpc::Scan::global: Likewise.
(Target_powerpc::Relocate::relocate): Delete shadowing r_sym.
A branch in a non-exec section that needs a stub can lead to this
assertion.
* powerpc.cc (Powerpc_relobj::stub_table): Return NULL rather
then asserting.
Adds a new option, defaulting to off, that allows a group of stubs to
serve multiple output sections. Prior to this patch powerpc gold
allowed this unconditionally, which is a little unsafe with clever
code that discards/reuses sections at runtime.
* options.h (--stub-group-multi): New PowerPC option.
* powerpc.cc (Stub_control): Add multi_os_ var and param
to constructor. Sort start_ var later. Comment State.
(Stub_control::can_add_to_stub_group): Heed multi_os_.
(Target_powerpc::group_sections): Update.
Gold attaches stubs to an existing section in contrast to ld.bfd which
inserts a new section for stubs. If we want stubs before branches,
then the stubs must be added to the previous section. Adding to the
previous section is a disaster if there is a large gap between the
previous section and the group.
PR gold/20878
* powerpc.cc (Stub_control): Replace stubs_always_before_branch_
with stubs_always_after_branch_, group_end_addr_ with
group_start_addr_.
(Stub_control::can_add_to_stub_group): Rewrite to suit scanning
sections by increasing address.
(Target_powerpc::group_sections): Scan that way. Delete corner
case.
* options.h (--stub-group-size): Update help string.
Some more debug output, and a little hardening.
* powerpc.cc (Stub_table_owner): Provide constructor.
(Powerpc_relobj::set_stub_table): Resize fill with -1.
(Target_powerpc::Branch_info::make_stub): Provide target debug
output on returning false.
This patch adds a little more debug output, and replaces two variables
with one, tracking current max group size by group_size_ rather than
by has14_.
* powerpc.cc (class Stub_control): Delete stub14_group_size_
and has14_. Add group_size_.
(Stub_control::can_add_to_stub_group): Adjust to suit. Print
debug info when switching to adding sections before stubs.
This patch rewrites the rather obscure can_add_to_stub_group, fixing
a problem with the handling of sections containing conditional
external branches. When a section group contains any such section,
the group size needs to be limited to a much smaller size than groups
with only non-conditional external branches.
PR 20523
* powerpc.cc (class Stub_control): Add has14_. Comment owner_.
(Stub_control::can_add_to_stub_group): Correct grouping of
sections containing 14-bit external branches. When returning
false, set state_ to reflect the fact that we have one section
for the next group. Rewrite most of function for clarity.
Add and expand comments.
(Target_powerpc::do_relax): Print stub group size retry in hex.
Alan Modra
Cary Coutant
Yuri Chornovian
James Clarke
H.J. Lu