Don't overload bfd_reloc_outofrange with what is really a domain error
(target at odd address), or an overflow.
PR 11290
* reloc.c (bfd_reloc_other): Correct comment.
* elf32-avr.c (avr_final_link_relocate): Return bfd_reloc_other
for unaligned reloc target values. Return bfd_reloc_overflow
when stubs are too far away and when R_AVR_LDS_STS_16,
R_AVR_PORT6, or R_AVR_PORT5 overflow.
(elf32_avr_relocate_section): Report more descriptive relocation
errors.
* bfd-in2.h: Regenerate.
On Windows, filename_cmp is case insensitive, but when cross compiling
with libraries that may contain members with uppercase file names, we
should keep those comparisons case insensitive when running the build
tools on other OSes too.
Also make the check for .def consistent with the other ones, fixing
out of bounds reads if file names are shorter than 4 characters.
There's been a long-standing bug in the arm backend where
target-specific directives did not correctly handle lines with
multiple statements. This patch fixes the issue for all the cases
I've been able to find.
It does result in a slight change in behaviour when errors are
encountered: where, previously,
.cpu arm6 bar
would result in the error "junk at end of line, first unrecognized
character is `b'", we now get "unknown cpu `arm6 bar'", which I think
is slightly more helpful anyway. Similar errors are generated for
other directives.
When debugging a certain class of GDB bug, I often end up wanting to
know what GDB thinks the frame-id is in a particular frame. It's
not too hard to pull this from some debug output, but I thought it
might be nice if there was a maintenance command that could tell us.
This commit adds 'maint print frame-id' which prints the frame-id of
the currently selected frame. You can also pass a frame level number
to find the frame-id for a specific frame.
There's a new test too.
Fix the bug that can not generate func@plt
when linking a undefined function with cmodel=medium.
Add testcase.
bfd/
* elfnn-loongarch.c
ld/testsuite/ld-loongarch-elf/
* cmodel-libjirl.dd
* cmodel.exp
* libjirl.s
I don't think it makes any sense for a SHT_RELR section to specify a
symbol table with sh_link. SHT_RELR relocations don't use symbols.
There is no real need to specify sh_info either, SHT_RELR is not for
relocatable object files. Anyway, fuzzers of course don't restrict
themselves to even half-sensible objects. So they found a hole in
objcopy using a non-alloc SHT_RELR in an ET_EXEC. In that case BFD
set up the SHT_RELR section as if it were a SHT_REL against the
sh_info target section. When it came to reading in the target section
relocs, the count was horribly wrong which caused a buffer overflow.
* elf.c (bfd_section_from_shdr <SHT_RELR>): Always just make a
normal section, don't treat it as a reloc section.
PR 29489
* dlltool.c (deterministic): New variable.
(gen_lib_file): If deterministic is true set the
BFD_DETERMINISTIC_OUTPUT flag.
(usage): Mention --deterministic-libraries and
--non-deterministic-libraries.
(long_options): Add new options.
(main): Parse new options.
* doc/binutils.texi: Document the new options.
* NEWS: Mention the new feature.
This implements target async for Windows. The basic idea is to have
the worker thread block in WaitForDebugEvent, then notify the event
loop when an event is seen. In a few situations, this blocking
behavior is undesirable, so the functions passed to do_synchronously
are changed to return a boolean indicating which behavior is needed.
On Windows, certain debugging APIs can only be called from the thread
that started (or attached) to the inferior. Also, there is no way on
Windows to wait for a debug event in addition to other events.
Therefore, in order to implement target async for Windows, gdb will
have to call some functions in a worker thread.
This patch implements the worker thread and moves the necessary
operations there. Target async isn't yet implemented, so this patch
does not cause any visible changes.
When using Ravenscar, gdb can crash if the user sets a breakpoint very
early in task startup. This happens because gdb thinks the runtime is
initialized, but in practice the particular task isn't sufficiently
initialized. This patch avoids the issue by turning an assertion into
an early return.
I tested this using the AdaCore internal test suite. I don't know how
to test Ravenscar using the FSF test suite.
* readelf.c (check_magic_number): New function. Checks the magic
bytes at the start of a file. If they are not the ELF format
magic values, then attempts to generate a helpful error message.
(process_file_header): Call check_magic_number.
Instructions that can load immediate support using constant
variable like ".equ var, 123 li.w/d resgister, var".
gas/
* config/loongarch-parse.y
* config/tc-loongarch.c
Add four testcases.One is a program using constant variable,
one test using label is unsupported, and another two test
almost instructions that can load immediate.
gas/
* testsuite/gas/loongarch/li.d
* testsuite/gas/loongarch/li.s
* testsuite/gas/loongarch/imm_ins_label-fail.d
* testsuite/gas/loongarch/imm_ins_label-fail.l
* testsuite/gas/loongarch/imm_ins_label-fail.s
* testsuite/gas/loongarch/imm_ins.d
* testsuite/gas/loongarch/imm_ins.s
* testsuite/gas/loongarch/imm_ins_32.d
* testsuite/gas/loongarch/imm_ins_32.s
I noticed that gdbpy_parse_register_id would assert if passed a Python
object of a type it was not expecting. The included test case shows
this crash. This patch fixes the problem and also changes
gdbpy_parse_register_id to be more "Python-like" -- it always ensures
the Python error is set when it fails, and the callers now simply
propagate the existing exception.
If symbols are provided by the caller of this function they are
passed on to bfd_get_relocated_section_contents. No surprises there.
It gets a little weird if they are not provided. In that case they
are read from the bfd by _bfd_generic_link_add_symbols, and global
symbols are added to the generic linker hash table. Global symbols
are not added to the linker hash table if symbols *are* provided. Now
the linker hash table symbols are not used by the generic
bfd_get_relocated_section_conents, and also not by most target
versions when called from bfd_simple_get_relocated_section_contents
except for symbols like "_gp". So it mostly doesn't matter whether
symbols are in the linker hash table, but it's odd that there is a
difference. We could always add them, but I'm inclined to think that
is unnecessary work so this patch always leaves them out.
Also, symbols are canonicalized and written into a malloc'd buffer.
The buffer isn't freed, see commit 8e16317ca5. I don't know whether
that matters any more, but in any case I can't see why we need another
copy of the symbols when _bfd_generic_link_read_symbols has already
cached symbols.
* simple.c (bfd_simple_get_relocated_section_contents): If not
provided, read symbols via bfd_generic_link_read_symbols. Do
not create another copy of symbols. Tidy failure exits.
Minor tidy of bfd_get_relocated_section_contents and
bfd_get_full_section_contents arguments.
Fixes this when cross-compiling from x86_64-linux
x86_64-w64-mingw32 +FAIL: compiling shared lib fastcall/stdcall
* testsuite/ld-pe/pe-run2-def.exp (test_direct2_link_dll_def):
Use CC_FOR_TARGET and CFLAGS_FOR_TARGET rather than CC and CFLAGS.
I noticed a couple of initialization functions that aren't really
needed, and that currently require explicit calls in gdb_init. This
patch removes these functions, simplifying gdb a little.
Regression tested on x86-64 Fedora 34.
As Luis pointed out here [1], the AArch64 variant of the test doesn't
work on systems that use PIE by default. For example, on this Debian
11:
$ make check TESTS="gdb.dwarf2/entry-value-typedef.exp"
gdb compile failed, /usr/bin/ld: /tmp/ccJE8ZSr.o: relocation R_AARCH64_ADR_PREL_PG_HI21 against symbol `_ZNSsD1Ev@@GLIBCXX_3.4' which may bind externally can not be used when making a shared object; recompile with -fPIC
/usr/bin/ld: /tmp/ccJE8ZSr.o(.text+0x38): unresolvable R_AARCH64_ADR_PREL_PG_HI21 relocation against symbol `_ZNSsD1Ev@@GLIBCXX_3.4'
This is because entry-value-typedef-aarch64.S was generated on an old
system that does not generate position-independent code by default, but
the system the test runs on tries to link the test executable as
position-independent. Fix this by regenerating the same binary on the
same system as the original one, but with -fPIE this time. Do the same
for the amd64 binary, although this one was already position-independent
so the generated code doesn't change.
With this patch applied, the test passes on the Debian 11 AArch64
system.
[1] https://sourceware.org/pipermail/gdb-patches/2022-August/191462.html
Change-Id: I68d55adaa56a7a3eddb0c13980b1a98b791f8144
Clang 15.0.0 enabled the warning for deprecated non-prototype functions
by default: https://reviews.llvm.org/D122895
Callfuncs.exp is impacted and won't run due to new warnings:
callfuncs.c:339:5: warning: a function declaration without a prototype is
deprecated in all versions of C and is not supported in C2x
[-Wdeprecated-non-prototype]
int t_float_values (float_arg1, float_arg2)
This patch disables those warnings with -Wno-deprecated-non-prototype.
Removing the test for deprecated syntax would also be an option. But I will
leave that up for others to decide/implement.
To cite gdb.exp:
Some C/C++ testcases unconditionally pass -Wno-foo as additional
options to disable some warning. That is OK with GCC, because
by design, GCC accepts any -Wno-foo option, even if it doesn't
support -Wfoo. Clang however warns about unknown -Wno-foo by
default, unless you pass -Wno-unknown-warning-option as well.
We do that here, so that individual testcases don't have to
worry about it.
This patch adds the same option that already exists for clang for icx and
adds the equivalent icc option.
According to LoongArch ELF ABI specification [1], variadic arguments
are passed in GARs in the same manner as named arguments. And after
a variadic argument has been passed on the stack, all future arguments
will also be passed on the stack, i.e., the last argument register may
be left unused due to the aligned register pair rule. long double data
tpye is passed in an aligned GAR pair, the first register in the pair
is even-numbered.
[1] https://loongson.github.io/LoongArch-Documentation/LoongArch-ELF-ABI-EN.html
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Bug 29374 shows this crash:
$ ./gdb -nx --data-directory=data-directory -q -batch -ex "catch throw" -ex r -ex bt a.out
...
/home/simark/src/binutils-gdb/gdb/../gdbsupport/array-view.h:217: internal-error: copy: Assertion `dest.size () == src.size ()' failed.
The backtrace is:
#0 internal_error (file=0x5555606504c0 "/home/simark/src/binutils-gdb/gdb/../gdbsupport/array-view.h", line=217, fmt=0x55556064b700 "%s: Assertion `%s' failed.") at /home/simark/src/binutils-gdb/gdbsupport/errors.cc:51
#1 0x000055555d41c0bb in gdb::copy<unsigned char const, unsigned char> (src=..., dest=...) at /home/simark/src/binutils-gdb/gdb/../gdbsupport/array-view.h:217
#2 0x000055555deef28c in dwarf_expr_context::fetch_result (this=0x7fffffffb830, type=0x621007a86830, subobj_type=0x621007a86830, subobj_offset=0, as_lval=false) at /home/simark/src/binutils-gdb/gdb/dwarf2/expr.c:1040
#3 0x000055555def0015 in dwarf_expr_context::evaluate (this=0x7fffffffb830, addr=0x62f00004313e "0", len=1, as_lval=false, per_cu=0x60b000069550, frame=0x621007c9e910, addr_info=0x0, type=0x621007a86830, subobj_type=0x621007a86830, subobj_offset=0) at /home/simark/src/binutils-gdb/gdb/dwarf2/expr.c:1091
#4 0x000055555e084327 in dwarf2_evaluate_loc_desc_full (type=0x621007a86830, frame=0x621007c9e910, data=0x62f00004313e "0", size=1, per_cu=0x60b000069550, per_objfile=0x613000006080, subobj_type=0x621007a86830, subobj_byte_offset=0, as_lval=false) at /home/simark/src/binutils-gdb/gdb/dwarf2/loc.c:1485
#5 0x000055555e0849e2 in dwarf2_evaluate_loc_desc (type=0x621007a86830, frame=0x621007c9e910, data=0x62f00004313e "0", size=1, per_cu=0x60b000069550, per_objfile=0x613000006080, as_lval=false) at /home/simark/src/binutils-gdb/gdb/dwarf2/loc.c:1529
#6 0x000055555e0828c6 in dwarf_entry_parameter_to_value (parameter=0x621007a96e58, deref_size=0x0, type=0x621007a86830, caller_frame=0x621007c9e910, per_cu=0x60b000069550, per_objfile=0x613000006080) at /home/simark/src/binutils-gdb/gdb/dwarf2/loc.c:1235
#7 0x000055555e082f55 in value_of_dwarf_reg_entry (type=0x621007a86890, frame=0x621007acc510, kind=CALL_SITE_PARAMETER_DWARF_REG, kind_u=...) at /home/simark/src/binutils-gdb/gdb/dwarf2/loc.c:1332
#8 0x000055555e083449 in value_of_dwarf_block_entry (type=0x621007a86890, frame=0x621007acc510, block=0x61e000033568 "T\004\205\001\240\004\004\243\001T\237\004\240\004\261\004\001T\004\261\004\304\005\004\243\001T\237\004\304\005\310\005\001T\004\310\005\311\005\004\243\001T\237", block_len=1) at /home/simark/src/binutils-gdb/gdb/dwarf2/loc.c:1365
#9 0x000055555e094d40 in loclist_read_variable_at_entry (symbol=0x621007a99bd0, frame=0x621007acc510) at /home/simark/src/binutils-gdb/gdb/dwarf2/loc.c:3889
#10 0x000055555f5192e0 in read_frame_arg (fp_opts=..., sym=0x621007a99bd0, frame=0x621007acc510, argp=0x7fffffffbf20, entryargp=0x7fffffffbf60) at /home/simark/src/binutils-gdb/gdb/stack.c:559
#11 0x000055555f51c352 in print_frame_args (fp_opts=..., func=0x621007a99ad0, frame=0x621007acc510, num=-1, stream=0x6030000bad90) at /home/simark/src/binutils-gdb/gdb/stack.c:887
#12 0x000055555f521919 in print_frame (fp_opts=..., frame=0x621007acc510, print_level=1, print_what=LOCATION, print_args=1, sal=...) at /home/simark/src/binutils-gdb/gdb/stack.c:1390
#13 0x000055555f51f22e in print_frame_info (fp_opts=..., frame=0x621007acc510, print_level=1, print_what=LOCATION, print_args=1, set_current_sal=0) at /home/simark/src/binutils-gdb/gdb/stack.c:1116
#14 0x000055555f526c6d in backtrace_command_1 (fp_opts=..., bt_opts=..., count_exp=0x0, from_tty=0) at /home/simark/src/binutils-gdb/gdb/stack.c:2079
#15 0x000055555f527ae5 in backtrace_command (arg=0x0, from_tty=0) at /home/simark/src/binutils-gdb/gdb/stack.c:2198
The problem is that the type that gets passed down to
dwarf_expr_context::fetch_result (the type of a variable of which we're
trying to read the entry value) is a typedef whose size has never been
computed yet (check_typedef has never been called on it). As we get in
the DWARF_VALUE_STACK case (line 1028 of dwarf2/expr.c), the `len`
variable is therefore set to 0, instead of the actual type length. We
then call allocate_value on subobj_type, which does call check_typedef,
so the length of the typedef gets filled in at that point. We end up
passing to the copy function a source array view of length 0 and a
target array view of length 4, and the assertion fails.
Fix this by calling check_typedef on both type and subobj_type at the
beginning of fetch_result.
I tried writing a test for this using the DWARF assembler, but I haven't
succeeded. It's possible that we need to get into this specific code
path (value_of_dwarf_reg_entry and all) to manage to get to
dwarf_expr_context::fetch_result with a typedef type that has never been
resolved. In all my attempts, the typedef would always be resolved
already, so the bug wouldn't show up.
As a fallback, I made a gdb.dwarf2 test with compiler-generated .S
files. I don't particularly like those, but I think it's better than no
test. The .cpp source code is the smallest reproducer I am able to make
from the reproducer given in the bug (thanks to Pedro for suggestions on
how to minimize it further than I had). Since I tested on both amd64
and aarch64, I added versions of the test for these two architectures.
Change-Id: I182733ad08e34df40d8bcc47af72c482fabf4900
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29374
As reported by Tom in https://sourceware.org/pipermail/gdb-patches/2022-August/191357.html,
the aarch64 prologue analyzer considers the adrp instruction in the
gdb.dwarf2/dw2-dir-file-name.exp testcase to be part of a prologue.
The function has no prologue though, and it only loads the volatile variable
from memory. GDB should not skip any instructions in this case.
Doing some archaeology, it seems handling for adr/adrp in prologues was
included with the original aarch64 port. It might've been an oversight.
In the particular case of gdb.dwarf2/dw2-dir-file-name.exp, the analyzer skips
a couple instructions and leaves us in a nice spot where the address to the
variable "v" is already in w0. But no prologues exists.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29481
This changes how bookmarks are allocated and stored, replacing a
linked list with a vector and removing some ALL_* iterator macros.
Regression tested on x86-64 Fedora 34.
It exercises a bug that GDB previously had where it would lose track of
some registers when the inferior changed its vector length.
It also checks that the vg register and the size of the z0-z31 registers
correctly reflect the new vector length.
When the inferior program changes the SVE length, GDB can stop tracking
some registers as it obtains the new gdbarch that corresponds to the
updated length:
Breakpoint 1, do_sve_ioctl_test () at sve-ioctls.c:44
44 res = prctl(PR_SVE_SET_VL, i, 0, 0, 0, 0);
(gdb) print i
$2 = 32
(gdb) info registers
⋮
[ snip registers x0 to x30 ]
⋮
sp 0xffffffffeff0 0xffffffffeff0
pc 0xaaaaaaaaa8ac 0xaaaaaaaaa8ac <do_sve_ioctl_test+112>
cpsr 0x60000000 [ EL=0 BTYPE=0 C Z ]
fpsr 0x0 0
fpcr 0x0 0
vg 0x8 8
tpidr 0xfffff7fcb320 0xfffff7fcb320
(gdb) next
45 if (res < 0) {
(gdb) info registers
⋮
[ snip registers x0 to x30 ]
⋮
sp 0xffffffffeff0 0xffffffffeff0
pc 0xaaaaaaaaa8cc 0xaaaaaaaaa8cc <do_sve_ioctl_test+144>
cpsr 0x200000 [ EL=0 BTYPE=0 SS ]
fpsr 0x0 0
fpcr 0x0 0
vg 0x4 4
(gdb)
Notice that register tpidr disappeared when vg (which holds the vector
length) changed from 8 to 4. The tpidr register is provided by the
org.gnu.gdb.aarch64.tls feature.
This happens because the code that searches for a new gdbarch to match the
new vector length in aarch64_linux_nat_target::thread_architecture doesn't
take into account the features present in the target description associated
with the previous gdbarch. This patch makes it do that.
Since the id member of struct gdbarch_info is now unused, it's removed.
Since commit b2d8657, having a per-interpreter event/command loop is not
possible anymore.
As Insight uses a GUI that has its own event loop, gdb and GUI event
loops have then to be "merged" (i.e.: work together). But this is
problematic as gdb_do_one_event is not aware of this alternate event
loop and thus may wait forever.
A solution is to delegate GUI events handling to the gdb events handler.
Insight uses Tck/Tk as GUI and the latter offers a "notifier" feature to
implement such a delegation. The Tcl notifier spec requires the event wait
function to support a timeout parameter. Unfortunately gdb_do_one_event
does not feature such a parameter.
This timeout cannot be implemented externally with a gdb timer, because
it would become an event by itself and thus can cause a legitimate event to
be missed if the timeout is 0.
Tcl implements "idle events" that are (internally) triggered only when no
other event is pending. For this reason, it can call the event wait function
with a 0 timeout quite often.
This patch implements a wait timeout to gdb_do_one_event. The initial
pending events monitoring is performed as before without the possibility
to enter a wait state. If no pending event has been found during this
phase, a timer is then created for the given timeout in order to re-use
the implemented timeout logic and the event wait is then performed.
This "internal" timer only limits the wait time and should never be triggered.
It is deleted upon gdb_do_one_event exit.
The new parameter defaults to "no timeout" (-1): as it is used by Insight
only, there is no need to update calls from the gdb source tree.
Emitting this warning for every insn, including ones having actual
errors, is annoying. Introduce a boolean variable to emit the warning
just once on the first insn after .arch may have changed the things, and
move the warning to output_insn(). (I didn't want to go as far as
checking whether the .arch actually turned off the i386 bit, but doing
so would be an option.)
