| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
|
|
For PDE, "recompiling with -fPIE" just makes no sense.
For PIE, "recompiling with -fPIE" makes sense for unresolvable absolute
relocs, but not unresolveable PC-relative relocs: if the reloc is
already PC-relative, the problem is not the reloc is PC-relative or
absolute, but the reloc is not applicable for external symbols.
If we hit an unresolvable reloc in PDE or an unresolvable PC-relative
reloc in PIE, it means the programmer has somehow wrongly instructed the
compiler to treat external symbols as local symbols. A misuse of
-mdirect-extern-access can cause the issue, so we can suggest
-mno-direct-extern-access. And in all cases (DSO/PIE/PDE) a mismatching
symbol visibility can also cause the issue, so we should also suggest to
check the visibility.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
weak symbols for static PIE
In a static PIE, undefined weak symbols should be just resolved to
runtime address 0, like those symbols with non-default visibility. This
was silently broken in all prior Binutils releases with "-static-pie
-mdirect-extern-access":
$ cat t.c
int x (void) __attribute__ ((weak));
int
main (void)
{
__builtin_printf("%p\n", x);
}
$ gcc t.c -static-pie -mdirect-extern-access
$ ./a.out
0x7ffff1d64000
Since commit 4cb77761d687 ("LoongArch: Check PC-relative relocations for
shared libraries), the situation has been improved: the linker errors
out instead of silently producing a wrong output file.
But logically, using -mdirect-extern-access for a static PIE perfectly
makes sense, and we should not prevent that even if the programmer uses
weak symbols. Linux kernel is such an example, and Linux < 6.10 now
fails to build with Binutils trunk. (The silent breakage with prior
Binutils releases was "benign" due to some blind luck.)
While since the 6.10 release Linux has removed those potentially
undefined weak symbols (due to performance issue), we still should
support weak symbols in -mdirect-extern-access -static-pie and unbreak
building old kernels.
Link: https://lore.kernel.org/loongarch/20241206085810.112341-1-chenhuacai@loongson.cn/
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
An undefined weak hidden/protect symbol should be resolved to runtime
address 0, but we were actually resolving it to link-time address 0. So
in PIE or DSO the runtime address would be incorrect.
Fix the issue by rewriting pcalau12i to lu12i.w, and pcaddi to addi.w.
The latter does not always work because the immediate field of addi.w is
narrower, report an error in the case the addend is too large.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
This is another case where the proper object_id can be read from
elf_backend_data.
|
|
hash_table_id can be set from elf_backend_data, now that all targets
have matching ELF_TARGET_ID and hash_table_init target_id.
|
|
Any _new_section_hook that is not itself called from another
_new_section_hook will always see used_by_bfd NULL. Remove those
NULL checks in such hooks, and tidy code a little.
|
|
As per the spec (Section 7.5.10, LoongArch Reference Manual Vol. 1),
LoongArch machines are not limited in page size choices, and currently
page sizes of 4KiB, 16KiB and 64KiB are supported by mainline Linux.
While 16KiB is the most common, the current BFD code says it is the
maximum; this is not correct, and as an effect, almost all existing
binaries are incompatible with a 64KiB kernel because the sections are
not sufficiently aligned, while being totally fine otherwise.
This is needlessly complicating integration testing [1].
This patch fixes the inconsistency, and also brings BFD behavior in line
with that of LLD [2].
[1] https://github.com/loongson-community/discussions/issues/47
[2] https://github.com/llvm/llvm-project/blob/llvmorg-19.1.0/lld/ELF/Arch/LoongArch.cpp#L174-L183
bfd/
* elfnn-loongarch.c (ELF_MAXPAGESIZE): Bump to 64KiB.
(ELF_MINPAGESIZE): Define as 4KiB.
(ELF_COMMONPAGESIZE): Define as 16KiB.
ld/
* testsuite/ld-loongarch-elf/64_pcrel.d: Update assertions after
changing the target max page size to 64KiB.
* testsuite/ld-loongarch-elf/data-got.d: Likewise.
* testsuite/ld-loongarch-elf/desc-relex.d: Likewise.
* testsuite/ld-loongarch-elf/relax-align-ignore-start.d: Likewise.
* testsuite/ld-loongarch-elf/tlsdesc_abs.d: Make the fuzzy match work
as intended by not checking exact instruction words.
* testsuite/ld-loongarch-elf/tlsdesc_extreme.d: Likewise.
Signed-off-by: WANG Xuerui <git@xen0n.name>
|
|
The undefweak symbol value of non-default visibility is 0 and does
not use plt entry, and will not be relocated in the relocate_secion
function. As a result, an infinite loop is generated because
bl %plt(sym) => bl 0.
Fix this by converting the call into a jump address 0.
|
|
There is no overflow check for the relaxation of pcalau12i+ld.d =>
pcalau12i+addi.d. For instruction sequences that can be relaxed,
they are directly relaxed to pcalau12i+addi.d. However, when the
relative distance between the symbol and the pc exceeds the 32-bit
range, the symbol value cannot be obtained correctly.
Adds an overflow check for the relaxation of pcalau12i+ld.d.
If it is found that the relaxation will overflow, it will not
be relaxed.
|
|
detail
|
|
|
|
Local absolute symbols are resolved at assembly stage and the symbol
value is placed in the relocation addend. But non-zero addend will
cause an assertion failure during linking.
Forces emission of relocations to defer resolution of local abs symbols
until link time.
bfd/
* elfnn-loongarch.c (loongarch_elf_relax_section): Determine
absolute symbols in advance to avoid ld crash.
gas/
* config/tc-loongarch.c (loongarch_force_relocation): New
function to force relocation.
* config/tc-loongarch.h (TC_FORCE_RELOCATION): New macros
to force relocation.
(loongarch_force_relocation): Function declaration.
* testsuite/gas/loongarch/localpic.d: New test.
* testsuite/gas/loongarch/localpic.s: New test.
|
|
Building shared libraries should not be allowed for PC-relative
relocations against external symbols.
Currently LoongArch has no corresponding checks and silently
generates wrong shared libraries.
However, In the first version of the medium cmodel, pcalau12i+jirl was
used for function calls, in which case PC-relative relocations were
allowed.
|
|
Add relaxation support for call36 that jump to PLT entry.
Add relaxation support for call36 with IFUNC symbol.
Add relaxation support for call36 that jump to undefweak symbol.
For undefweak symbol, it can always be relaxed if it have no PLT entry.
Because we set the address of undefweak symbol without PLT entry to PC
like relocate_section.
|
|
The symbol value is only calculated when the relocation can be relaxed.
|
|
LoongArch: Add elfNN_loongarch_mkobject to initialize LoongArch tdata.
|
|
LoongArch: Not append rela for absolute symbol
Use la.global to get absolute symbol like la.abs.
la.global put address of a global symbol into a got
entry and append a rela for it, which will be used
to relocate by dynamic linker. Dynamic linker should
not relocate for got entry of absolute symbol as it
stores symval not symbol's address.
|
|
LoongArch: Add macros to get opcode and register of instructions appropriately
Currently, we get opcode of an instruction by manipulate the binary with
it's mask, it's a bit of a pain. Now a macro is defined to do this and a
macro to get the RD and RJ registers which is applicable to most instructions
of LoongArch are added.
|
|
Commit "b67a17aa7c0c478a" modified the logic of allocating dynamic
relocation space for TLS GD/IE, but only modified the logic of
generation dynamic relocations for TLS GD/IE in ABI v2.00. When
linking an object file of ABI v1.00 with bfd ld of ABI v2.00, it
will cause an assertion failure.
Modified the dynamic relocation generation logic of TLS GD/IE
in ABI v1.00 to be consistent with ABI v2.00.
|
|
If the symbol defined by PROVIDE in the link script is not in SECTION,
the symbol is placed in the ABS section. The linker considers that
symbols in the ABS section do not need to calculate PC relative offsets.
Symbols in ABS sections should calculate PC relative offsets normally
based on relocations.
|
|
Due to the way BFD implements DT_RELR as a part of relaxation, if in a
relax trip size_relative_relocs has changed the layout, when
relax_section runs only the vma of the section being relaxed is
guaranteed to be updated. Then bad thing can happen. For example, when
linking an auxilary program _freeze_module in the Python 3.12.4 building
system (with PGO + LTO), before sizing the .relr.dyn section, the vma of
.text is 30560 and the vma of .rodata is 2350944; in the final
executable the vma of .text is 36704 and the vma of .rodata is 2357024.
The vma increase is expected because .relr.dyn is squashed somewhere
before .text. But size_relative_relocs may see the state in which .text
is at 36704 but .rodata "is" still at 2350944. Thus the distance
between .text and .rodata can be under-estimated and overflowing
R_LARCH_PCREL20_S2 reloc can be created.
To avoid this issue, if size_relative_relocs is updating the size of
.relr.dyn, just supress the normal relaxation in this relax trip. In
this situation size_relative_relocs should have been demending the next
relax trip, so the normal relaxation can happen in the next trip.
I tried to make a reduced test case but failed.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
In the DT_RELR implementation I missed a code path emiting relative
reloc entries. Then the already packed relative reloc entries will be
(unnecessarily) pushed into .rela.dyn but we've not allocated the space
for them, triggering an assertion failure.
Unfortunately I failed to notice the issue until profiled bootstrapping
GCC with LTO and -Wl,-z,pack-relative-relocs. The failure can be easily
triggered by linking a "hello world" program with -fprofile-generate and
LTO:
$ PATH=$HOME/ld-test:$PATH gcc hw.c -fprofile-generate -Wl,-z,pack-relative-relocs -flto
/home/xry111/git-repos/binutils-build/TEST/ld: BFD (GNU Binutils) 2.43.50.20240802 assertion fail ../../binutils-gdb/bfd/elfnn-loongarch.c:2628
/home/xry111/git-repos/binutils-build/TEST/ld: BFD (GNU Binutils) 2.43.50.20240802 assertion fail ../../binutils-gdb/bfd/elfnn-loongarch.c:2628
collect2: error: ld returned 1 exit status
And the reduced test case is just incredibly simple (included in the
patch) so it seems I'm just stupid enough to fail to detect it before.
Let's fix it now anyway.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
Some tests started passing with commit 3a83f0342e54. However,
supporting a changed ld output format is not so simple, and the change
to the loongarch_elf_hash_table macro needs further changes to the
rest of the code. It is true that some uses of
loongarch_elf_hash_table do not need to check the type of the hash
table, but others like loongarch_elf_relax_section do need to check.
bfd_relax_section is called in lang_size_sections using the input bfd,
not the output bfd. If the input bfd may be of different type to the
output, then the hash table type must be checked before accessing
elements of the hash table. This patch corrects
loongarch_elf_relax_section. I haven't checked all the uses of the
hash table throughout the loongarch backend.
bfd/
* elfnn-loongarch.c (loongarch_elf_relax_section): Don't relax
unless the hash table is loongarch_elf_link_hash_table.
Move variable declarations. Formatting.
ld/
* testsuite/ld-elf/pr21884.d: Don't xfail loongarach.
* testsuite/ld-unique/pr21529.d: Likewise.
|
|
In binary output format, loongarch_elf_hash_table return NULL and result
in segment fault.
When ld output binary file, it seems that elf related functions should
not be called. But loongarch_elf_relax_section be called and
loongarch_elf_hash_table cause segment fault.
Just redefined loongarch_elf_hash_table and always return
link_info->hash.
The tests of binutils, glibc and gcc is ok.
0 loongarch_elf_relax_section ()
1 0x000055555557ab28 in lang_size_sections_1 ()
2 0x000055555557a16c in lang_size_sections_1 ()
3 0x000055555557b0a8 in one_lang_size_sections_pass ()
4 0x000055555557b478 in lang_size_sections ()
5 0x000055555557e65c in lang_relax_sections ()
6 0x000055555559f9c8 in ldelf_map_segments ()
7 0x000055555559783c in gldelf64loongarch_after_allocation ()
8 0x000055555558dac0 in ldemul_after_allocation ()
9 0x000055555557f6c0 in lang_process ()
10 0x0000555555585314 in main ()
|
|
The absolute symbol should be resolved to const when link to dso or exe.
Alloc dynamic relocs will cause extra space and R_LARCH_NONE finally.
|
|
The logic is same as a71d87680110 ("aarch64: Add DT_RELR support").
As LoongArch does not have -z dynamic-undefined-weak, we don't need to
consider UNDEFWEAK_NO_DYNAMIC_RELOC.
The linker relaxation adds another layer of complexity. When we delete
bytes in a section during relaxation, we need to fix up the offset in
the to-be-packed relative relocations against this section.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
On LoongArch there is no reason to treat STV_PROTECTED STT_FUNC symbols
as preemptible. See the comment above LARCH_REF_LOCAL for detailed
explanation.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
With a simple test case:
.globl ifunc
.globl ifunc_hidden
.hidden ifunc_hidden
.type ifunc, %gnu_indirect_function
.type ifunc_hidden, %gnu_indirect_function
.text
.align 2
ifunc: ret
ifunc_hidden: ret
test:
bl ifunc
bl ifunc_hidden
"ld -shared" produces a shared object with one R_LARCH_NONE (instead of
R_LARCH_JUMP_SLOT as we expect) to relocate the GOT entry of "ifunc".
It's because the indices in .plt and .rela.plt mismatches for
STV_DEFAULT STT_IFUNC symbols when another PLT entry exists for a
STV_HIDDEN STT_IFUNC symbol, and such a mismatch breaks the logic of
loongarch_elf_finish_dynamic_symbol. Fix the issue by reordering .plt
so the indices no longer mismatch.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
We were converting R_LARCH_32 to R_LARCH_RELATIVE for ELFCLASS64:
$ cat t.s
.data
x:
.4byte x
.4byte 0xdeadbeef
$ as/as-new t.s -o t.o
$ ld/ld-new -shared t.o
$ objdump -R
a.out: file format elf64-loongarch
DYNAMIC RELOCATION RECORDS
OFFSET TYPE VALUE
00000000000001a8 R_LARCH_RELATIVE *ABS*+0x00000000000001a8
But this is just wrong: at runtime the dynamic linker will run
*(uintptr *)&x += load_address, clobbering the next 4 bytes of data
("0xdeadbeef" in the example).
If we keep the R_LARCH_32 reloc as-is in ELFCLASS64, it'll be rejected
by the Glibc dynamic linker anyway. And it does not make too much sense
to modify Glibc to support it. So we can just reject it like x86_64:
relocation R_X86_64_32 against `.data' can not be used when making a
shared object; recompile with -fPIC
or RISC-V:
relocation R_RISCV_32 against non-absolute symbol `a local symbol'
can not be used in RV64 when making a shared object
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
old object files
R_LARCH_SOP_PUSH_ABSOLUTE with -fPIC was heavily used in the era of gas-2.38.
We do not check this relocation to prevent broken links with old object
files.
|
|
As the comment in the code says, TLS_IE needs only one dynamic reloc.
But commit b67a17aa7c0c ("LoongArch: Fix the issue of excessive
relocation generated by GD and IE") has incorrectly allocated the space
for two dynamic relocs, causing libc.so to contain 8 R_LARCH_NONE.
Adjust tlsdesc-dso.d for the offset changes and add two tests to ensure
there are no R_LARCH_NONE with TLS.
Signed-off-by: Xi Ruoyao <xry111@xry111.site>
|
|
R_LARCH_ALIGN (psABI v2.30) requires a symbol index. The symbol is only
created at the first time to handle alignment directive. This means that
all other sections may use this symbol. If the section of this symbol is
discarded, there may be problems. Search it in its own section.
Remove elf_backend_data.is_rela_normal() function added at commit daeda14191c.
Co-authored-by: Jinyang He <hejinyang@loongson.cn>
Reported-by: WANG Xuerui <git@xen0n.name>
Link: https://lore.kernel.org/loongarch/2abbb633-a10e-71cc-a5e1-4d9e39074066@loongson.cn/T/#t
|
|
ld -z separate-code let .text and .rodata in two different but read only
segment. If the symbol and pc in two segment, the offset from pc to
symbol need to consider segment alignment.
Add a function 'loongarch_two_sections_in_same_segment' to determine
whether two sections are in the same segment.
|
|
|
|
Absolute address symbols cannot be used with -shared.
We output more information to the user than just BFD_ASSETR.
|
|
When scanning relocations and determining whether TLS type transition is
possible, it will try to obtain the symbol got type. If the symbol got
type record has not yet been allocated space and initialized, it will
cause ld to crash. So when uninitialized, the symbol is set to GOT_UNKNOWN.
|
|
If we generate an object file using an assembler with the new
relocations added, and then linking those files with an older
linker, the link will still complete and the linked file will
be generated.
In this case we should report an error instead of continuing
the linking process.
|
|
Move .got .got.plt before .data so .got can be protected with -zrelro.
And the first two entries of .got.plt (_dl_runtime_resolve and link map)
are placed within the relro region.
|
|
This largely mechanical patch is preparation for a followup patch.
For quite some time I've thought that it would be useful to call
elf_backend_size_dynamic_sections even when no dynamic objects are
seen by the linker. That's what this patch does, with some renaming.
There are no functional changes to the linker, just a move of the
dynobj test in bfd_elf_size_dynamic_sections to target backend
functions, replacing the asserts/aborts already there. No doubt some
of the current always_size_sections functions could be moved to
size_dynamic_sections but I haven't made that change.
Because both hooks are now always called, I have renamed
always_size_sections to early_size_sections and size_dynamic_sections
to late_size_sections. I condisdered calling late_size_sections plain
size_sections, since this is the usual target dynamic section sizing
hook, but decided that searching the sources for "size_sections" would
then hit early_size_sections and other functions.
|
|
Currently, whether GD and IE generate dynamic relocation is
determined by SYMBOL_REFERENCES_LOCAL and bfd_link_executable.
This results in dynamic relocations still being generated in some
situations where dynamic relocations are not necessary (such as
the undefined weak symbol in static links).
We use RLARCH_TLS_GD_IE_NEED_DYN_RELOC macros to determine whether
GD/IE needs dynamic relocation. If GD/IE requires dynamic relocation,
set need_reloc to true and indx to be a dynamic index.
At the same time, some test cases were modified to use regular
expression matching instead of complete disassembly matching.
|
|
To represent the first and third expression of .align, R_LARCH_ALIGN need to
associate with a symbol. We define a local symbol for R_LARCH_AGLIN.
But if the section of the local symbol is discarded, it may result in
a undefined symbol error.
Instead, we use the section name symbols, and this does not need to
add extra symbols.
During partial linking (ld -r), if the symbol associated with a relocation is
STT_SECTION type, the addend of relocation needs to add the section output
offset. We prevent it for R_LARCH_ALIGN.
The elf_backend_data.rela_normal only can set all relocations of a target to
rela_normal. Add a new function is_rela_normal to elf_backend_data, it can
set part of relocations to rela_normal.
|
|
This relaxation is effective for both macro instructions (call36, tail36)
and explicit relocation instructions (pcaddu18i + jirl).
call36 f -> bl f
R_LARCH_CALL36 -> R_LARCH_B26
tail36 $t0, f -> b f
R_LARCH_CALL36 -> R_LARCH_B26
|
|
When tls_sec is NULL, we should not access the virtual address
of tls_sec.
|
|
This modification mainly changes the timing of type transition,
adds relaxation to the old LE instruction sequence, and fixes
bugs in extreme code models.
We strictly distinguish between type transition and relaxation.
Type transition is from one type to another, while relaxation
is the removal of instructions under the same TLS type. Detailed
instructions are as follows:
1. For type transition, only the normal code model of DESC/IE
does type transition, and each relocation is accompanied by a
RELAX relocation. Neither abs nor extreme will do type transition,
and no RELAX relocation will be generated.
The extra instructions when DESC transitions to other TLS types
will be deleted during the type transition.
2. Implemented relaxation for the old LE instruction sequence.
The first two instructions of LE's 32-bit and 64-bit models
use the same relocations and cannot be distinguished based on
relocations. Therefore, for LE's instruction sequence, any code
model will try to relax.
3. Some function names have been adjusted to facilitate understanding,
parameters have been adjusted, and unused macros have been deleted.
|
|
--unresolved-symbols=ignore-all
Remove duplicated check when producing executable files that reference external symbols
defined in other files. RELOC_FOR_GLOBAL_SYMBOL will check it.
Testcase is:
resolv.c:
int main(int argc, char *argv[]) {
return argc;
}
t.c:
extern const struct my_struct ms1;
static const struct my_struct *ms = &ms1;
t.h:
typedef struct my_struct {
char *str;
int i;
} my_struct;
Compiling and linking command with:
gcc t.c -c ; gcc resolv.c -c
gcc resolv.o t.o -o resolv -Wl,--unresolved-symbols=ignore-all
Got error as:
~/install/usr/bin/ld: t.o:(.data.rel+0x0): undefined reference to `ms1'
collect2: error: ld returned 1 exit status
|
|
The relsec size is still increased although sec is discarded, which
cause a lot of unused space allocated. Avoid size increased if sec
was discarded.
bfd/ChangeLog:
* bfd/elfnn-loongarch.c: (allocate_dynrelocs): Do not increase
sreloc size when discarded_section.
ld/ChangeLog:
* ld/testsuite/ld-loongarch-elf/ld-loongarch-elf.exp: Add test.
* ld/testsuite/ld-loongarch-elf/pie_discard.d: New test.
* ld/testsuite/ld-loongarch-elf/pie_discard.s: New test.
* ld/testsuite/ld-loongarch-elf/pie_discard.t: New test.
|
|
The old code works because R_LARCH_RELAX has no symbol index. It causes
'(rel + 1)->r_info == R_LARCH_RELAX' is 1 and ELFNN_R_TYPE (1) is 1.
|
|
Add support for TLS LE symbol with addend, such as:
lu12i.w $t0, %le_hi20(a + 0x8)
ori $t0, $t0, %le_lo12(a + 0x8)
|
|
TLS LE is exclusively for executables, while DF_STATIC_TLS is for DLLs.
DF_STATIC_TLS should only be set for TLS IE (and when it's DLL), not LE.
|
