This provides a linker generated __tls_get_addr_desc wrapper function
preserving registers around a __tls_get_addr call. The idea being to
support __tls_get_addr_desc without requiring a glibc update.
bfd/
* elf64-ppc.c (struct ppc_link_hash_table): Add tga_group.
(ppc64_elf_archive_symbol_lookup): Extract __tls_get_addr_opt for
__tls_get_addr_desc.
(ppc64_elf_size_stubs): Add section for linker generated
__tls_get_addr_desc wrapper function. Loop at least once if
generating this function.
(emit_tga_desc, emit_tga_desc_eh_frame): New functions.
(ppc64_elf_build_stubs): Generate __tls_get_addr_desc.
ld/
* testsuite/ld-powerpc/tlsdesc3.d,
* testsuite/ld-powerpc/tlsdesc3.wf,
* testsuite/ld-powerpc/tlsdesc4.d,
* testsuite/ld-powerpc/tlsdesc4.s,
* testsuite/ld-powerpc/tlsdesc4.wf: New tests.
* testsuite/ld-powerpc/powerpc.exp: Run them.
This implements register saving and restoring in the __tls_get_addr
call stub, so that when glibc supports the optimized tls call stub gcc
can generate code that assumes only r0, r12 and of course r3 are
changed on a __tls_get_addr call. When gcc expects __tls_get_addr
calls to preserve registers the call will be to __tls_get_addr_desc,
which will be translated by the linker to a call to __tls_get_addr_opt.
bfd/
* elf64-ppc.h (struct ppc64_elf_params): Add no_tls_get_addr_regsave.
* elf64-ppc.c (struct ppc_link_hash_table): Add tga_desc and
tga_desc_fd.
(is_tls_get_addr): Match tga_desc and tga_desc_df too.
(STDU_R1_0R1, ADDI_R1_R1): Define.
(tls_get_addr_prologue, tls_get_addr_epilogue): New functions.
(ppc64_elf_tls_setup): Set up tga_desc and tga_desc_fd. Indirect
tga_desc_fd to opt_fd, and tga_desc to opt. Set
no_tls_get_addr_regsave.
(branch_reloc_hash_match): Add hash3 and hash4.
(ppc64_elf_tls_optimize): Handle tga_desc_fd and tga_desc too.
(ppc64_elf_size_dynamic_sections): Likewise.
(ppc64_elf_relocate_section): Likewise.
(plt_stub_size, build_plt_stub): Likewise. Size regsave
__tls_get_addr stub.
(build_tls_get_addr_stub): Build regsave __tls_get_addr stub and
eh_frame.
(ppc_size_one_stub): Handle tga_desc_fd and tga_desc too. Size
eh_frame for regsave __tls_get_addr.
gas/
* config/tc-ppc.c (parse_tls_arg): Handle tls arg for
__tls_get_addr_desc and __tls_get_addr_opt.
ld/
* emultempl/ppc64elf.em (ppc64_opt, PARSE_AND_LIST_LONGOPTS),
(PARSE_AND_LIST_OPTIONS, PARSE_AND_LIST_ARGS_CASES): Support
--tls-get-addr-regsave and --no-tls-get-addr-regsave.
(params): Init new field.
* ld.texi (--tls-get-addr-regsave, --no-tls-get-addr-regsave):
Document.
* testsuite/ld-powerpc/tlsdesc.s,
* testsuite/ld-powerpc/tlsdesc.d,
* testsuite/ld-powerpc/tlsdesc.wf,
* testsuite/ld-powerpc/tlsdesc2.d,
* testsuite/ld-powerpc/tlsdesc2.wf,
* testsuite/ld-powerpc/tlsexenors.d,
* testsuite/ld-powerpc/tlsexenors.r,
* testsuite/ld-powerpc/tlsexers.d,
* testsuite/ld-powerpc/tlsexers.r,
* testsuite/ld-powerpc/tlsexetocnors.d,
* testsuite/ld-powerpc/tlsexetocrs.d,
* testsuite/ld-powerpc/tlsexetocrs.r,
* testsuite/ld-powerpc/tlsopt6.d,
* testsuite/ld-powerpc/tlsopt6.wf: New.
* testsuite/ld-powerpc/powerpc.exp: Run new tests.
When linking with --no-tls-optimize the linker doesn't generate a call
or long branch stub to __tls_get_addr in some circumstances, giving:
relocation truncated to fit: R_PPC64_REL24 against symbol `__tls_get_addr'
* elf64-ppc.c (ppc64_elf_size_stubs): Correct condition under
which __tls_get_addr calls will be eliminated.
This modifies the special __tls_get_addr stub that checks for a
tlsdesc style __tls_index entry and returns early. Not using r11
isn't much benefit at the moment but a followup patch will preserve
regs around the first call to __tls_get_addr when the __tls_index
entry isn't yet set up for an early return.
bfd/
* elf64-ppc.c (LD_R11_0R3, CMPDI_R11_0, STD_R11_0R1, LD_R11_0R1),
(MTLR_R11): Don't define.
(LD_R0_0R3, CMPDI_R0_0): Define.
(build_tls_get_addr_stub): Don't use r11 in stub.
ld/
* testsuite/ld-powerpc/tlsexe.d: Match new __tls_get_addr stub.
* testsuite/ld-powerpc/tlsexeno.d: Likewise.
* testsuite/ld-powerpc/tlsexetoc.d: Likewise.
* testsuite/ld-powerpc/tlsexetocno.d: Likewise.
* testsuite/ld-powerpc/tlsopt5.d: Likewise.
Function symbols of course don't normally want .dynbss copies but
with some old versions of gcc they are needed to copy the function
descriptor. This patch restricts the cases where they are useful to
compilers using dot-symbols, and enables the warning regardless of
whether a PLT entry is emitted in the executable. PLTs in shared
libraries are affected by a .dynbss copy in the executable.
bfd/
PR 25384
* elf64-ppc.c (ELIMINATE_COPY_RELOCS): Update comment.
(ppc64_elf_adjust_dynamic_symbol): Don't allow .dynbss copies
of function symbols unless dot symbols are present. Do warn
whenever one is created, regardles of whether a PLT entry is
also emitted for the function symbol.
ld/
* testsuite/ld-powerpc/ambiguousv1b.d: Adjust expected output.
* testsuite/ld-powerpc/funref.s: Align func_tab.
* testsuite/ld-powerpc/funref2.s: Likewise.
* testsuite/ld-powerpc/funv1.s: Add dot symbols.
I thought the fuzzers were really going overboard by defining
VMS_DEBUG but that wasn't the case. VMS_DEBUG is defined by
default. Let's not do that, and fix the segv as well.
* vms.h (VMS_DEBUG): Define as 0.
* vms-alpha.c (image_write): Move debug output after bounds check.
Tidy bounds check.
(_bfd_vms_slurp_eihd): Warning fix.
(_bfd_vms_slurp_etir): Init variables to avoid bogus warnings.
It's always a bad idea to perform arithmetic on an unknown value read
from an object file before comparing against bounds. Code like the
following attempting to bounds check "len", a 64-bit value, isn't
effective because the pointer arithmetic ignores the high 32 bits when
compiled for a 32-bit host.
READ_LEB128 (len, p, end);
if (p + len < p || p + len > end)
goto error_return;
Instead, perform any arithmetic on known values where we don't need to
worry about overflows:
READ_LEB128 (len, p, end);
if (len > (size_t) (end - p))
goto error_return;
I'll note that this check does do things the right way:
READ_LEB128 (symcount, p, end);
/* Sanity check: each symbol has at least two bytes. */
if (symcount > payload_size / 2)
return FALSE;
"symcount * 2 > payload_size" would be wrong since the multiply could
overflow.
* wasm-module.c (wasm_scan_name_function_section): Formatting.
Delete asect name check. Move asect NULL check to wasm_object_p.
Correct bounds check of sizes against end. Replace uses of
bfd_zalloc with bfd_alloc, zeroing only necessary bytes. Use
just one bfd_release.
(wasm_scan): Don't use malloc/strdup for section names,
bfd_alloc instead. Simplify code prefixing section name.
Formatting. Don't attempt to free memory here..
(wasm_object_p): ..do so here. Formatting.
With static PIE linking undefined weak symbols are resolved to 0, so no
dynamic relocation is needed for them. The UNDEFWEAK_NO_DYNAMIC_RELOC
macro was introduced so this case can be handled easily, but it was not
applied consistently in the first attempt to fix ld/22269 for arm:
commit 95b03e4ad68e7a90f5096b47df595636344b783a
arm: Check UNDEFWEAK_NO_DYNAMIC_RELOC
This patch fixes spurious relative relocs in static PIE binaries against
GOT entries created for undefined weak symbols on arm*-*, this fixes
FAIL: pr22269-1 (static pie undefined weak)
bfd/ChangeLog:
PR ld/22269
* elf32-arm.c (elf32_arm_final_link_relocate): Use
UNDEFWEAK_NO_DYNAMIC_RELOC.
(allocate_dynrelocs_for_symbol): Likewise.
This changes the fix to PR 25210 by removing the ELF class change.
As it turns out the correct change was only the change in compress.c.
Everything else is unneeded and setting the elf class is making the linker
behave very oddly under LTO. The first stub is correctly written out but for
the rest the suddenly don't have a pointer to the stub section anymore.
This caused SPEC to fail as the program would branch to the stub and it wouldn't
be filled in.
Committed to master under the trivial rule as this is partially reverting a previous commit.
bfd/ChangeLog:
PR 25210
* elfnn-aarch64.c (_bfd_aarch64_create_stub_section): Remove elfclass.
Since 1993-11-05 git commit c188b0bec3b, bfd_check_format has failed
if any of the target object_p functions returns false with any error
but bfd_error_wrong_format. That's just weird. There is really no
reason why coff_real_object_p should be fixed to only return that
error instead of numerous other possible errors. Even an out of
memory condition for one target doesn't necessarily mean other targets
can't match, assuming the failing target nicely returns all memory it
might have used.
* format.c (bfd_check_format_matches): Ignore bfd_error on target
match failures. Don't init to bfd_error_wrong_format before
calling _bfd_check_format.
BFD is not supposed to exit or abort on anything the user can do.
* vms-alpha.c (_bfd_vms_push, _bfd_vms_pop): Return pass/fail
status rather than exiting on stack overflow or underflow.
(_bfd_vms_slurp_etir): Adjust to suit.
Fixes lots of places the fuzzers are going to find, and the one they
already hit.
* mach-o.c (bfd_mach_o_read_dylinker): Don't read past end of
command. Check name offset is within command.
(bfd_mach_o_read_dylib, bfd_mach_o_read_prebound_dylib),
(bfd_mach_o_read_prebind_cksum, bfd_mach_o_read_twolevel_hints),
(bfd_mach_o_read_fvmlib, bfd_mach_o_read_dysymtab),
(bfd_mach_o_read_symtab, bfd_mach_o_read_uuid),
(bfd_mach_o_read_linkedit, bfd_mach_o_read_str),
(bfd_mach_o_read_dyld_info, bfd_mach_o_read_version_min),
(bfd_mach_o_read_encryption_info, bfd_mach_o_read_source_version),
(bfd_mach_o_read_encryption_info_64, bfd_mach_o_read_main),
(bfd_mach_o_read_note, bfd_mach_o_read_build_version),
(bfd_mach_o_read_segment): Similarly.
(bfd_mach_o_read_thread): Properly bound check thread struct.
Don't repeat checks on second loop.
(bfd_mach_o_read_command): Fail on invalid command length.
It's a little tricky. We can release any memory back when we have a
match failure, but after a match success which we might want to
preserve for later use the high water mark must change to that of the
matched bfd.
* format.c (bfd_check_format_matches): Add preserve_match.
Save initial bfd state in "preserve", matched bfd state in
"preserve_match". Save just the first match. Release
bfd_alloc memory. Restore and finish preserved state as
appropriate on all function exit paths.
object_p functions cannot allocate memory by malloc and not free it
before returning. Even a successful target match may not be the best
match. If a match isn't used then those malloc'd blocks won't be
freed.
* coffgen.c (coff_real_object_p): Free malloc'd memory on target
match too.
PR 25307
(bfd_pef_parse_function_stubs): Correct the test that ensures that
there is enough data remaining in the code buffer before
attempting to read a function stub.
