Add assertions to ensure we don't access an array_view out of bounds.
Enable these assertions only when _GLIBCXX_DEBUG is set, as we did for
gdb::optional.
Change-Id: Iffaee38252405073735ed123c8e57fde6b2c6be3
I wanted to write a warning that included two target_pid_to_str calls,
like this:
warning (_("Blabla %s, blabla %s"),
target_pid_to_str (ptid1),
target_pid_to_str (ptid2));
This doesn't work, because target_pid_to_str stores its result in a
static buffer, so my message would show twice the same ptid. Change
target_pid_to_str to return an std::string to avoid this. I don't think
we save much by using a static buffer, but it is more error-prone.
Change-Id: Ie3f649627686b84930529cc5c7c691ccf5d36dc2
With test-case gdb.mi/mi-var-cp.exp I run into this duplicate:
...
PASS: gdb.mi/mi-var-cp.exp: run to mi-var-cp.cc:104 (set breakpoint)
PASS: gdb.mi/mi-var-cp.exp: create varobj for s
PASS: gdb.mi/mi-var-cp.exp: create varobj for s
DUPLICATE: gdb.mi/mi-var-cp.exp: create varobj for s
...
This is due to a duplicate test name here:
...
$ cat -n gdb/testsuite/gdb.mi/mi-var-cp.cc
...
100 int reference_to_struct ()
101 {
102 /*: BEGIN: reference_to_struct :*/
103 S s = {7, 8};
104 S& r = s;
105 /*:
106 mi_create_varobj S s "create varobj for s"
107 mi_create_varobj R r "create varobj for s"
...
Fix this by using "create varobj for r" instead.
Tested on x86_64-linux.
ctf_type_visit (used, among other things, by the type dumping code) was
aborting when it saw a nonrepresentable type anywhere: even a single
structure member with a nonrepresentable type caused an abort with
ECTF_NONREPRESENTABLE. This is not useful behaviour, given that the
abort comes from a type-resolution we are only doing in order to
determine whether the type is a structure or union. We know
nonrepresentable types can't be either, so handle that case and
pass the nonrepresentable type down.
(The added test verifies that the dumper now handles this case and
prints nonrepresentable structure members as it already does
nonrepresentable top-level types, rather than skipping the whole
structure -- or, without the previous commit, skipping the whole types
section.)
ld/ChangeLog
2021-10-25 Nick Alcock <nick.alcock@oracle.com>
* testsuite/ld-ctf/nonrepresentable-member.*: New test.
libctf/ChangeLog
2021-10-25 Nick Alcock <nick.alcock@oracle.com>
* ctf-types.c (ctf_type_rvisit): Handle nonrepresentable types.
If dumping of a single type fails, we obviously can't dump it; but just
as obviously this doesn't make the other types in the types section
invalid or undumpable. So we should not propagate errors seen when
type-dumping, but rather ignore them and carry on, so we dump as many
types as we can (leaving out the ones we can't grok).
libctf/ChangeLog
2021-10-25 Nick Alcock <nick.alcock@oracle.com>
* ctf-dump.c (ctf_dump_type): Do not abort on error.
This option has been present since the very early days of the
development of libctf as part of binutils, and it shows. Back in the
earliest days, I thought we might handle ambiguous types by introducing
new ELF sections on the fly named things like .ctf.foo.c for ambiguous
types found only in foo.c, etc. This turned out to be a terrible idea,
so we moved to using a CTF archive in the .ctf section which contained
all the CTF dictionaries -- but the --ctf-parent option in objdump and
readelf was never adjusted, and lingered as a mechanism to specify CTF
parent dictionaries in sections other than .ctf, even though the linker
has no way to produce parent dictionaries in different sections from
their children, libctf's ctf_open can't handle such split-up
parent/child dicts, and they are never found in the wild, emitted by GNU
ld or by any known third-party linking tool.
Meanwhile, the actually-useful ctf_link feature (albeit not used by ld)
which lets you remap the names of CTF archive members (so you can end up
with a parent archive member named something other than ".ctf", still
contained with all its children in a single .ctf section) had no support
in objdump or readelf: there was no way to tell them that these members
were parents, so all the types in the associated child dicts always
appeared corrupted, referencing nonexistent types from a parent objdump
couldn't find.
So adjust --ctf-parent so that rather than taking a section name it
takes a member name instead (if not specified, the name is ".ctf", which
is what GNU ld emits). Because the option was always useless before
now, this is expected to have no backward-compatibility implications.
As part of this, we have to slightly adjust the code which skips the
archive member name if redundant: right now it skips it if it's ".ctf",
on the assumption that this name will almost always be at the start
of the objdump output and thus we'll end up with a shared dump
and then smaller, headed dumps for the per-TU child dicts; but if
the parent name has been changed, that won't be true any more.
So change the rules to "members named .ctf which appear first in the
first have their member name skipped". Since we now need to count
members, move from ctf_archive_iter (for which passing in extra
parameters requires defining a new struct and is clumsy) to
ctf_archive_next, allowing us to just *call* dump_ctf_archive_member and
maintain a member count in the obvious way. In the process we fix a
tiny difference between readelf and objdump: if a ctf_dump ever failed,
readelf skipped every later member, while objdump tried to keep going as
much as it could. For a dumping tool the former is clearly preferable.
binutils/ChangeLog
2021-10-25 Nick Alcock <nick.alcock@oracle.com>
* objdump.c (usage): --ctf-parent now takes a name, not a section.
(dump_ctf): Don't open a separate section; use the parent_name in
ctf_dict_open instead. Use ctf_archive_next, not ctf_archive_iter,
so we can pass down a member count.
(dump_ctf_archive_member): Add the member count; don't return
anything. Import parents into children no matter what the
parent's name, while still avoiding displaying the header for the
common parent name of ".ctf".
* readelf.c (usage): Adjust similarly.
(dump_section_as_ctf): Likewise.
(dump_ctf_archive_member): Likewise. Never stop iterating over
archive members, even if ctf_dump of one member fails.
* doc/ctf.options.texi: Adjust.
Sometimes the investigation of a fuzzing bug report leads into areas
you'd rather not go. In this instance by the time I'd figured out the
real cause was a target variant that had never been properly supported
in binutils, the time needed to fix it was less than the time needed
to rip it out.
* coffcode.h (coff_set_alignment_hook): Call bfd_coff_swap_reloc_in
not coff_swap_reloc_in.
(coff_slurp_reloc_table): Likewise. Don't use RELOC type.
(ticoff0_swap_table): Use coff_swap_reloc_v0_out and
coff_swap_reloc_v0_in.
* coffswap.h (coff_swap_reloc_v0_in, coff_swap_reloc_v0_out): New.
* coff-tic54x.c (tic54x_lookup_howto): Don't abort.
* coffgen.c (coff_get_normalized_symtab): Use PTR_ADD.
* bfd-in.h (PTR_ADD, NPTR_ADD): Avoid warnings when passing an
expression.
* bfd-in2.h: Regenerate.
PR 21813
* mach-o-arm.c (bfd_mach_o_arm_canonicalize_one_reloc): Sanity
check PAIR reloc in other branch of condition as was done for
PR21813. Formatting. Delete debug printf.
Unaligned load/store instructions on aligned memory or register are as
fast as aligned load/store instructions on modern Intel processors. Add
a command-line option, -muse-unaligned-vector-move, to x86 assembler to
encode encode aligned vector load/store instructions as unaligned
vector load/store instructions.
* NEWS: Mention -muse-unaligned-vector-move.
* config/tc-i386.c (use_unaligned_vector_move): New.
(encode_with_unaligned_vector_move): Likewise.
(md_assemble): Call encode_with_unaligned_vector_move for
-muse-unaligned-vector-move.
(OPTION_MUSE_UNALIGNED_VECTOR_MOVE): New.
(md_longopts): Add -muse-unaligned-vector-move.
(md_parse_option): Handle -muse-unaligned-vector-move.
(md_show_usage): Add -muse-unaligned-vector-move.
* doc/c-i386.texi: Document -muse-unaligned-vector-move.
* testsuite/gas/i386/i386.exp: Run unaligned-vector-move and
x86-64-unaligned-vector-move.
* testsuite/gas/i386/unaligned-vector-move.d: New file.
* testsuite/gas/i386/unaligned-vector-move.s: Likewise.
* testsuite/gas/i386/x86-64-unaligned-vector-move.d: Likewise.
This adds some missing code to the 'uninstall' targets in gdb and
gdbserver. It also changes gdb's uninstall target so that it no
longer tries to remove any man page -- this is already done (and more
correctly) by doc/Makefile.in.
I tested this with 'make install' followed by 'make uninstall', then
examining the install tree for regular files. Only the 'dir' file
remains, but this appears to just be how 'install-info' is intended to
work.
This removes a number of unused variables from gdbserver's Makefile.
I found these while working on the subsequent patches, and figured it
would be cleaner to have a separate patch for the deletions.
On openSUSE Tumbleweed with glibc-debuginfo installed I get:
...
(gdb) PASS: gdb.threads/linux-dp.exp: continue to breakpoint: thread 5's print
where^M
#0 print_philosopher (n=3, left=33 '!', right=33 '!') at linux-dp.c:105^M
#1 0x0000000000401628 in philosopher (data=0x40537c) at linux-dp.c:148^M
#2 0x00007ffff7d56b37 in start_thread (arg=<optimized out>) \
at pthread_create.c:435^M
#3 0x00007ffff7ddb640 in clone3 () \
at ../sysdeps/unix/sysv/linux/x86_64/clone3.S:81^M
(gdb) PASS: gdb.threads/linux-dp.exp: first thread-specific breakpoint hit
...
while without debuginfo installed I get instead:
...
(gdb) PASS: gdb.threads/linux-dp.exp: continue to breakpoint: thread 5's print
where^M
#0 print_philosopher (n=3, left=33 '!', right=33 '!') at linux-dp.c:105^M
#1 0x0000000000401628 in philosopher (data=0x40537c) at linux-dp.c:148^M
#2 0x00007ffff7d56b37 in start_thread () from /lib64/libc.so.6^M
#3 0x00007ffff7ddb640 in clone3 () from /lib64/libc.so.6^M
(gdb) FAIL: gdb.threads/linux-dp.exp: first thread-specific breakpoint hit
...
The problem is that the regexp used:
...
"\(from .*libpthread\|at pthread_create\|in pthread_create\)"
...
expects the 'from' part to match libpthread, but in glibc 2.34 libpthread has
been merged into libc.
Fix this by updating the regexp.
Tested on x86_64-linux.
In a future commit I'm going to be creating gdb.Membuf objects from a
new file within gdb/python/py*.c. Currently all gdb.Membuf objects
are created directly within infpy_read_memory (as a result of calling
gdb.Inferior.read_memory()).
Initially I split out the Membuf creation code into a new function,
and left the new function in gdb/python/py-inferior.c, however, it
felt a little random that the Membuf creation code should live with
the inferior handling code.
So, then I moved all of the Membuf related code out into a new file,
gdb/python/py-membuf.c, the interface is gdbpy_buffer_to_membuf, which
wraps an array of bytes into a gdb.Membuf object.
Most of the code is moved directly from py-inferior.c with only minor
tweaks to layout and replacing NULL with nullptr, hence, I've left the
copyright date on py-membuf.c as 2009-2021 to match py-inferior.c.
Currently, the only user of this code is still py-inferior.c, but in
later commits this will change.
There should be no user visible changes after this commit.
Add a new function to the Python API, gdb.architecture_names(). This
function returns a list containing all of the supported architecture
names within the current build of GDB.
The values returned in this list are all of the possible values that
can be returned from gdb.Architecture.name().
The disassemble_info structure has four callbacks, we have three of
them as static member functions within gdb_disassembler, the fourth is
just a global static function.
However, this fourth callback, is still only used from the
disassemble_info struct, so there's no real reason for its special
handling.
This commit makes fprintf_disasm a static method within
gdb_disassembler.
There should be no user visible changes after this commit.
Consider the the pcgp-relax-02 testcase,
.text
.globl _start
_start:
.L1: auipc a0, %pcrel_hi(data_a)
.L2: auipc a1, %pcrel_hi(data_b)
addi a0, a0, %pcrel_lo(.L1)
addi a1, a1, %pcrel_lo(.L2)
.data
.word 0x0
.globl data_a
data_a:
.word 0x1
.section .rodata
.globl data_b
data_b:
.word 0x2
If the first auipc is deleted, but we are still building the pcgp
table (connect the high and low pcrel relocations), then there is
an aliasing issue that we need some way to disambiguate which of
the two symbols we are targeting. Therefore, Palmer thought of a
way to use R_RISCV_DELETE to split this into two phases, so we
could resolve the addresses before creating the ambiguities.
This patch just add the ld testcase for the above case, in case we
have changed something but break this.
ld/
* testsuite/ld-riscv-elf/ld-riscv-elf.exp: Renamed pcgp-relax
to pcgp-relax-01, and added pcgp-relax-02.
* testsuite/ld-riscv-elf/pcgp-relax-01.d: Renmaed from pcgp-relax.
* testsuite/ld-riscv-elf/pcgp-relax-01.s: Likewise.
* testsuite/ld-riscv-elf/pcgp-relax-02.d: New testcase.
* testsuite/ld-riscv-elf/pcgp-relax-02.s: Likewise.
Commit abd20cb637 and
ebdcad3fdd introduced additional
complexity into the paths run by the RISC-V relaxation pass in order to
resolve the issue of accurately keeping track of pcrel_hi and pcrel_lo
pairs. The first commit split up relaxation of these relocs into a pass
which occurred after other relaxations in order to prevent the situation
where bytes were deleted in between a pcrel_lo/pcrel_hi pair, inhibiting
our ability to find the corresponding pcrel_hi relocation from the
address attached to the pcrel_lo.
Since the relaxation was split into two passes the 'again' parameter
could not be used to perform the entire relaxation process again and so
the second commit added a way to restart ldelf_map_segments, thus
starting the whole process again.
Unfortunately this process could not account for the fact that we were
not finished with the relaxation process so in some cases - such as the
case where code would not fit in a memory region before the
R_RISCV_ALIGN relocation was relaxed - sanity checks in generic code
would fail.
This patch fixes all three of these concerns by reverting back to a
system of having only one target relax pass but updating entries in the
table of pcrel_hi/pcrel_lo relocs every time any bytes are deleted. Thus
we can keep track of the pairs accurately, and we can use the 'again'
parameter to restart the entire target relax pass, behaving in the way
that generic code expects. Unfortunately we must still have an
additional pass to delay deleting AUIPC bytes to avoid ambiguity between
pcrel_hi relocs stored in the table after deletion. This pass can only
be run once so we may potentially miss out on relaxation opportunities
but this is likely to be rare.
https://sourceware.org/bugzilla/show_bug.cgi?id=28410
bfd/
* elfnn-riscv.c (riscv_elf_link_hash_table): Removed restart_relax.
(riscv_elf_link_hash_table_create): Updated.
(riscv_relax_delete_bytes): Moved after the riscv_update_pcgp_relocs.
Update the pcgp_relocs table whenever bytes are deleted.
(riscv_update_pcgp_relocs): Add function to update the section
offset of pcrel_hi and pcrel_lo, and also update the symbol value
of pcrel_hi.
(_bfd_riscv_relax_call): Need to update the pcgp_relocs table
when deleting codes.
(_bfd_riscv_relax_lui): Likewise.
(_bfd_riscv_relax_tls_le): Likewise.
(_bfd_riscv_relax_align): Once we've handled an R_RISCV_ALIGN,
we can't relax anything else, so set the sec->sec_flg0 to true.
Besides, we don't need to update the pcgp_relocs table at this
stage, so just pass NULL pointer as the pcgp_relocs table for
riscv_relax_delete_bytes.
(_bfd_riscv_relax_section): Use only one pass for all target
relaxations.
(_bfd_riscv_relax_delete): Likewise, we don't need to update
the pcgp_relocs table at this stage, and don't need to set
the `again' since restart_relax mechanism is abandoned.
(bfd_elfNN_riscv_restart_relax_sections): Removed.
(_bfd_riscv_relax_section): Updated.
* elfxx-riscv.h (bfd_elf32_riscv_restart_relax_sections): Removed.
(bfd_elf64_riscv_restart_relax_sections): Likewise.
ld/
* emultempl/riscvelf.em: Revert restart_relax changes and set
relax_pass to 3.
* testsuite/ld-riscv-elf/align-small-region.d: New testcase.
* testsuite/ld-riscv-elf/align-small-region.ld: Likewise.
* testsuite/ld-riscv-elf/align-small-region.s: Likewise.
* testsuite/ld-riscv-elf/restart-relax.d: Removed sine the
restart_relax mechanism is abandoned.
* testsuite/ld-riscv-elf/restart-relax.s: Likewise.
* testsuite/ld-riscv-elf/ld-riscv-elf.exp: Updated.
Commit 183be22290 ("gdb, gdbserver: make target_waitstatus safe")
broke the remote-sim.c build. In fact, it does some wrong changes,
result of a bad sed invocation.
Fix it by adjusting the code to the new target_waitstatus API.
Change-Id: I3236ff7ef7681fc29215f68be210ff4263760e91
I stumbled on a bug caused by the fact that a code path read
target_waitstatus::value::sig (expecting it to contain a gdb_signal
value) while target_waitstatus::kind was TARGET_WAITKIND_FORKED. This
meant that the active union field was in fact
target_waitstatus::value::related_pid, and contained a ptid. The read
signal value was therefore garbage, and that caused GDB to crash soon
after. Or, since that GDB was built with ubsan, this nice error
message:
/home/simark/src/binutils-gdb/gdb/linux-nat.c:1271:12: runtime error: load of value 2686365, which is not a valid value for type 'gdb_signal'
Despite being a large-ish change, I think it would be nice to make
target_waitstatus safe against that kind of bug. As already done
elsewhere (e.g. dynamic_prop), validate that the type of value read from
the union matches what is supposed to be the active field.
- Make the kind and value of target_waitstatus private.
- Make the kind initialized to TARGET_WAITKIND_IGNORE on
target_waitstatus construction. This is what most users appear to do
explicitly.
- Add setters, one for each kind. Each setter takes as a parameter the
data associated to that kind, if any. This makes it impossible to
forget to attach the associated data.
- Add getters, one for each associated data type. Each getter
validates that the data type fetched by the user matches the wait
status kind.
- Change "integer" to "exit_status", "related_pid" to "child_ptid",
just because that's more precise terminology.
- Fix all users.
That last point is semi-mechanical. There are a lot of obvious changes,
but some less obvious ones. For example, it's not possible to set the
kind at some point and the associated data later, as some users did.
But in any case, the intent of the code should not change in this patch.
This was tested on x86-64 Linux (unix, native-gdbserver and
native-extended-gdbserver boards). It was built-tested on x86-64
FreeBSD, NetBSD, MinGW and macOS. The rest of the changes to native
files was done as a best effort. If I forgot any place to update in
these files, it should be easy to fix (unless the change happens to
reveal an actual bug).
Change-Id: I0ae967df1ff6e28de78abbe3ac9b4b2ff4ad03b7
Add a constructor to initialize the waitstatus members. Initialize the
others in the class directly.
Change-Id: I10f885eb33adfae86e3c97b1e135335b540d7442
Add a constructor and a destructor. The constructor takes care of the
initialization that happened in add_thread, while the destructor takes
care of the freeing that happened in free_one_thread. This is needed to
make target_waitstatus non-POD, as thread_info contains a member of that
type.
Change-Id: I1db321b4de9dd233ede0d5c101950f1d6f1d13b7
Since the two locations which check the debug arch are the same code currently, it is
a good idea to factor it out to a new function and just use that function from
aarch64_linux_get_debug_reg_capacity. This is also the first step to support
ARMv8.4 debug arch.
In commit 81e6b8eb20 "Make tui-winsource not use breakpoint_chain", a loop
body was transformed into a lambda function body:
...
- for (bp = breakpoint_chain;
- bp != NULL;
- bp = bp->next)
+ iterate_over_breakpoints ([&] (breakpoint *bp) -> bool
...
and consequently:
- a continue was replaced by a return, and
- a final return was added.
Then in commit 240edef62f "gdb: remove iterate_over_breakpoints function", we
transformed back to a loop body:
...
- iterate_over_breakpoints ([&] (breakpoint *bp) -> bool
+ for (breakpoint *bp : all_breakpoints ())
...
but without reverting the changes that introduced the two returns.
Consequently, breakpoints no longer show up in the tui source window.
Fix this by reverting the changes that introduced the two returns.
Build on x86_64-linux, tested with all .exp test-cases that contain
tuiterm_env.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=28483
The test expect the runto_main to stop at the first line of the function.
Depending on the optimization level, gdb may stop in the prolog or after
the prolog at the first line. To ensure the test stops at the first line
of main, have it explicitly stop at a break point on the first line of the
function.
On PowerPC, the test passes when compiled with no optimization but fails
with all levels of optimization due to gdb stopping in the prolog.
The "add accessors for field (and call site) location" patch caused a
gdb crash when running the internal AdaCore testsuite. This turned
out to be a latent bug in ada-lang.c.
The immediate cause of the bug is that find_struct_field
unconditionally uses TYPE_FIELD_BITPOS. This causes an assert for a
dynamic type.
This patch fixes the problem by doing two things. First, it changes
find_struct_field to use a dummy value for the field offset in the
situation where the offset is not actually needed by the caller. This
works because the offset isn't used in any other way -- only as a
result.
Second, this patch assures that calls to find_struct_field use a
resolved type when the offset is needed. For
value_tag_from_contents_and_address, this is done by resolving the
type explicitly. In ada_value_struct_elt, this is done by passing
nullptr for the out parameters when they are not needed (the second
call in this function already uses a resolved type).
Note that, while we believe the parent field probably can't occur at a
variable offset, the patch still updates this code path, just in case.
I've updated an existing test case to reproduce the crash.
I'm checking this in.
ldelf.c: In function 'ldelf_after_open':
ldelf.c:1049:43: warning: the comparison will always evaluate as 'true' for the address of 'elf_header' will never be NULL [-Waddress]
1049 | && elf_tdata (abfd)->elf_header != NULL
| ^~
In file included from ldelf.c:37:
../bfd/elf-bfd.h:1957:21: note: 'elf_header' declared here
1957 | Elf_Internal_Ehdr elf_header[1]; /* Actual data, but ref like ptr */
* ldelf.c (ldelf_after_open): Remove useless elf_header test.
Mainline gcc:
readelf.c: In function 'find_section':
readelf.c:349:8: error: the comparison will always evaluate as 'true' for the pointer operand in 'filedata->section_headers + (sizetype)((long unsigned int)i * 80)' must not be NULL [-Werror=address]
349 | ((X) != NULL \
| ^~
readelf.c:761:9: note: in expansion of macro 'SECTION_NAME_VALID'
761 | if (SECTION_NAME_VALID (filedata->section_headers + i)
| ^~~~~~~~~~~~~~~~~~
This will likely be fixed in gcc, but inline functions are nicer than
macros.
* readelf.c (SECTION_NAME, SECTION_NAME_VALID),
(SECTION_NAME_PRINT, VALID_SYMBOL_NAME, VALID_DYNAMIC_NAME),
(GET_DYNAMIC_NAME): Delete. Replace with..
(section_name, section_name_valid, section_name_print),
(valid_symbol_name, valid_dynamic_name, get_dynamic_name): ..these
new inline functions. Update use throughout file.
My previous PR27625 patch had a problem or two. For one, the error
"__tls_get_addr call lacks marker reloc" on processing some calls
before hitting a call without markers typically isn't seen. Instead a
gold assertion fails. Either way it would be a hard error, which
triggers on a file contained in libphobos.a when running the gcc
testsuite. A warning isn't even appropriate since the call involved
is one built by hand without any of the arg setup relocations that
might result in linker optimisation.
So this patch reverts most of commit 0af4fcc25d, instead entirely
ignoring the problem of mis-optimising old-style __tls_get_addr calls
without marker relocs. We can't handle them gracefully without
another pass over relocations before decisions are made about GOT
entries in Scan::global or Scan::local. That seems too costly, just
to link object files from 2009. What's more, there doesn't seem to be
any way to allow the libphobos explicit __tls_get_addr call, but not
old TLS sequences without marker relocs. Examining instructions
before the __tls_get_addr call is out of the question: program flow
might reach the call via a branch. Putting an R_PPC64_TLSGD marker
with zero sym on the call might be a solution, but current linkers
will then merrily optimise away the call!
PR gold/27625
* powerpc.cc (Powerpc_relobj): Delete no_tls_marker_, tls_marker_,
and tls_opt_error_ variables and accessors. Remove all uses.
