git commit d1e304bc27 was aimed at stopping uninitialised memory
access to the index_offset array. Unfortunately that patch resulted
in a different array being uninitialised for all instructions with
more than two arguments.
* ns32k-dis.c (print_insn_arg): Update comment.
(print_insn_ns32k): Reduce size of index_offset array, and
initialize, passing -1 to print_insn_arg for args that are not
an index. Don't exit arg loop early. Abort on bad arg number.
Fix comma at end of enumerator list seen with -std=c++98.
* plugin-api.h (enum ld_plugin_symbol_type): Remove
comma after last value of an enum.
* lto-symtab.h (enum gcc_plugin_symbol_type): Likewise.
dwarf2_evaluate_property should not modify its "addr_stack"
parameter's contents. This patch makes this part of the API, by
marking it const.
gdb/ChangeLog
2020-03-20 Tom Tromey <tromey@adacore.com>
* dwarf2/loc.h (dwarf2_evaluate_property): Make "addr_stack"
const.
* dwarf2/loc.c (dwarf2_evaluate_property): Make "addr_stack"
const.
When running test-case gdb.threads/omp-par-scope.exp, I get this XPASS:
...
XPASS: gdb.threads/omp-par-scope.exp: nested_parallel: outer_threads: \
outer stop: get valueof "num"
...
for test:
...
set thread_num [get_valueof "" "num" "unknown"]
...
The intention of the test is to get the value of local variable num, which
has been set to:
...
int num = omp_get_thread_num ();
...
but the actually printed value is 'num':
...
(gdb) print num^M
$76 = num^M
...
This is due to the fact that num is missing in the locals, so instead we find
the enum member 'num' of enum expression_operator in glibc/intl/plural-exp.h.
Fix this by getting the value using a new proc get_local_valueof, which uses
the "info locals" commands to get the value.
Tested on x86_64-linux, with gcc 7.5.0 (where the test xfails) and gcc
10.0.1 (where the test passes).
I stumbled on this snippet in nat/gdb_ptrace.h:
/* Some systems, in particular DEC OSF/1, Digital Unix, Compaq Tru64
or whatever it's called these days, don't provide a prototype for
ptrace. Provide one to silence compiler warnings. */
#ifndef HAVE_DECL_PTRACE
extern PTRACE_TYPE_RET ptrace();
#endif
I believe this is unnecessary today and should be removed. First, the
comment only mentions OSes we don't support (and to be honest, I had
never even heard of).
But most importantly, in C++, a declaration with empty parenthesis
declares a function that accepts no arguments, unlike in C. So if this
declaration was really used, GDB wouldn't build, since all ptrace call
sites pass some arguments. Since we haven't heard anything about this
causing some build failures since we have transitioned to C++, I
conclude that it's not used.
This patch removes it as well as the corresponding configure check.
gdb/ChangeLog:
* ptrace.m4: Don't check for ptrace declaration.
* config.in: Re-generate.
* configure: Re-generate.
* nat/gdb_ptrace.h: Don't declare ptrace if HAVE_DECL_PTRACE is
not defined.
gdbserver/ChangeLog:
* config.in: Re-generate.
* configure: Re-generate.
gdbsupport/ChangeLog:
* config.in: Re-generate.
* configure: Re-generate.
Linux returns long from ptrace(2) and BSDs int.
gdb/ChangeLog:
* amd64-bsd-nat.c (gdb_ptrace): Change return type from `int' to
`PTRACE_TYPE_RET'.
* i386-bsd-nat.c (gdb_ptrace): Likewise.
* sparc-nat.c (gdb_ptrace): Likewise.
* x86-bsd-nat.c (gdb_ptrace): Likewise.
The "restrict" patch added some asserts to c-exp.y, but one spot was
copy-pasted and referred to the wrong table. This was pointed out by
-fsanitize=address. This patch fixes the bug.
gdb/ChangeLog
2020-03-20 Tom Tromey <tromey@adacore.com>
* c-exp.y (lex_one_token): Fix assert.
I configured with -fsanitize=address and built gdb. linux-tdep.c and
ada-tasks.c failed to build due to some stringop-truncation errors,
e.g.:
In function ‘char* strncpy(char*, const char*, size_t)’,
inlined from ‘int linux_fill_prpsinfo(elf_internal_linux_prpsinfo*)’ at ../../binutils-gdb/gdb/linux-tdep.c:1742:11,
inlined from ‘char* linux_make_corefile_notes(gdbarch*, bfd*, int*)’ at ../../binutils-gdb/gdb/linux-tdep.c:1878:27:
/usr/include/bits/string_fortified.h:106:34: error: ‘char* __builtin_strncpy(char*, const char*, long unsigned int)’ specified bound 81 equals destination size [-Werror=stringop-truncation]
This patch fixes the problem by using "sizeof - 1" in the call to
strndup, as recommended in the GCC manual. This doesn't make a
difference here because the next line, in all cases, sets the final
element to '\0' anyway.
gdb/ChangeLog
2020-03-20 Tom Tromey <tromey@adacore.com>
* ada-tasks.c (read_atcb): Use smaller length in strncpy call.
* linux-tdep.c (linux_fill_prpsinfo): Use smaller length in
strncpy call.
Andrew Burgess pointed out on irc that "maint info line-table" doesn't
properly align the table headers. This patch fixes the problem by
switching the table to use ui-out.
This required a small tweak to one test case, as ui-out will pad a
field using spaces, even at the end of a line.
gdb/ChangeLog
2020-03-20 Tom Tromey <tromey@adacore.com>
* symmisc.c (maintenance_print_one_line_table): Use ui_out.
gdb/testsuite/ChangeLog
2020-03-20 Tom Tromey <tromey@adacore.com>
* gdb.dwarf2/dw2-ranges-base.exp: Update regular expressions.
The removal of val_print caused a regression in the Ada code. In one
scenario, a variant type would not be properly printed, because the
address of a component was lost. This patch fixes the bug by changing
this API to be value-based. This is cleaner and fixes the bug as a
side effect.
gdb/ChangeLog
2020-03-20 Tom Tromey <tromey@adacore.com>
* ada-valprint.c (print_variant_part): Remove parameters; switch
to value-based API.
(print_field_values): Likewise.
(ada_val_print_struct_union): Likewise.
(ada_value_print_1): Update.
gdb/testsuite/ChangeLog
2020-03-20 Tom Tromey <tromey@adacore.com>
* gdb.ada/sub_variant/subv.adb: New file.
* gdb.ada/sub_variant.exp: New file.
gdb/ChangeLog:
* ppc-nbsd-nat.c (ppc_nbsd_nat_target): Inherit from
nbsd_nat_target instead of inf_ptrace_target.
* ppc-nbsd-nat.c: Include "nbsd-nat.h", as we are now using
nbsd_nat_target.
NetBSD ptrace(2) accepts thread id (LWP) as the 4th argument for threads.
gdb/ChangeLog:
* hppa-nbsd-nat.c (fetch_registers): New variable lwp and pass
it to the ptrace call.
* (store_registers): Likewise.
When running test-cases gdb.threads/step-over-lands-on-breakpoint.exp and
gdb.threads/step-over-trips-on-watchpoint.exp with target board
unix/-flto/-O0/-flto-partition=none/-ffat-lto-objects, we run into timeouts
due not being able to set a breakpoint and then trying to continue to that
breakpoint.
In total, we run into 186 timeouts, which roughly corresponds to half an hour:
...
$ grep "FAIL.*(timeout)" gdb.sum \
| awk '{print $2}' \
| sort \
| uniq -c
66 gdb.threads/step-over-lands-on-breakpoint.exp:
120 gdb.threads/step-over-trips-on-watchpoint.exp:
...
Fix this by bailing out if the first break fails.
Tested on x86_64-linux, both with native and with target board mentioned above.
gdb/testsuite/ChangeLog:
2020-03-20 Tom de Vries <tdevries@suse.de>
* gdb.threads/step-over-lands-on-breakpoint.exp (do_test): Bail out if
first break fails.
* gdb.threads/step-over-trips-on-watchpoint.exp: (do_test): Same.
NetBSD ptrace(2) accepts thread id (LWP) as the 4th argument for threads.
gdb/ChangeLog:
* ppc-nbsd-nat.c (fetch_registers): New variable lwp and pass
it to the ptrace call.
* (store_registers): Likewise.
Since LTO plugin may generate more than one ltrans.o file from one input
IR object as LTO wrapper ignores -flto-partition=none:
lto-wrapper.c:608:
604 /* Drop arguments that we want to take from the link line. */
605 case OPT_flto_:
606 case OPT_flto:
607 case OPT_flto_partition_:
608 continue;
the LTO wrapper approach is not only slow but also unreliable. Since
the LTO plugin API has been extended to add LDPT_ADD_SYMBOLS_V2 with
symbol type and section kind, we can use LDPT_ADD_SYMBOLS_V2 to get
symbol type, instead of invoking the LTO wrapper.
PR binutils/25640
* plugin.c (plugin_list_entry): Add has_symbol_type.
(add_symbols_v2): New function.
(bfd_plugin_open_input): Don't invoke LTO wrapper if LTO plugin
provides symbol type.
(try_load_plugin): Add LDPT_ADD_SYMBOLS_V2.
(bfd_plugin_canonicalize_symtab): Use LTO plugin symbol type if
available.
* nds32-dis.c (print_insn_nds32): Remove unnecessary casts.
Initialize parts of buffer not written when handling a possible
2-byte insn at end of section. Don't attempt decoding of such
an insn by the 4-byte machinery.
We shouldn't really decode a 2-byte left-over at the end of a section
as if the section contains two more bytes of zeros. Not that it
matters very much, but this patch tidies the corner case.
* ppc-dis.c (print_insn_powerpc): Only clear needed bytes of
partially filled buffer. Prevent lookup of 4-byte insns when
only VLE 2-byte insns are possible due to section size. Print
".word" rather than ".long" for 2-byte leftovers.
Function pointers in elfNN_bed that are initialized by elfxx-target.h
to non-zero values generally don't need a non-NULL test before calling
them. Targets don't set a non-NULL function to NULL. The one
exception being elfnn-ia64.c and that exception is removed here.
* elf.c (_bfd_elf_setup_sections): Don't test known non-NULL
backend functions for NULL before calling.
(copy_special_section_fields, _bfd_elf_copy_private_bfd_data),
(bfd_section_from_shdr, assign_section_numbers): Likewise.
* elfcode.h (elf_write_relocs, elf_slurp_reloc_table): Likewise.
* elfnn-ia64.c (ignore_errors): New function.
(elf_backend_link_order_error_handler): Redefine as ignore_errors.
Unlike most other Operating Systems, NetBSD tracks both pid and lwp.
The process id on NetBSD is stored always in the pid field of ptid.
gdb/ChangeLog:
* inf-ptrace.h: Disable get_ptrace_pid on NetBSD.
* inf-ptrace.c: Likewise.
* (gdb_ptrace): Add.
* (inf_ptrace_target::resume): Update.
* (inf_ptrace_target::xfer_partial): Likewise.
* (inf_ptrace_peek_poke): Change argument `pid' to `ptid'.
* (inf_ptrace_peek_poke): Update.
PR 25676
bfd * dwarf2.c (struct varinfo): Add unit_offset field to record the
location of the varinfo in the unit's debug info data. Change the
type of the stack field to a boolean.
(lookup_var_by_offset): New function. Returns the varinfo
structure for the variable described at the given offset in the
unit's debug info.
(scan_unit_for_symbols): Add support for variables which have the
DW_AT_specification attribute.
binutils* testsuite/binutils-all/dw4.s: New test source file.
* testsuite/binutils-all/nm.exp: Run the new test.
I was doing some SVE tests on system QEMU and noticed quite a few failures
related to inferior function calls. Any attempt to do an inferior function
call would result in the following:
Unable to set VG register.: Success.
This happens because, after an inferior function call, GDB attempts to restore
the regcache state and updates the SVE register in order. Since the Z registers
show up before the VG register, VG is still INVALID by the time the first Z
register is being updated. So when executing the following code in
aarch64_sve_set_vq:
if (reg_buf->get_register_status (AARCH64_SVE_VG_REGNUM) != REG_VALID)
return false;
By returning false, we signal something is wrong, then we get to this:
/* First store vector length to the thread. This is done first to ensure the
ptrace buffers read from the kernel are the correct size. */
if (!aarch64_sve_set_vq (tid, regcache))
perror_with_name (_("Unable to set VG register."));
Ideally we'd always have a valid VG before attempting to set the Z registers,
but in this case the ordering of registers doesn't make that possible.
I considered reordering the registers to put VG before the Z registers, like
the DWARF numbering, but that would break backwards compatibility with
existing implementations. Also, the Z register numbering is pinned to the V
registers, and adding VG before Z would create a gap for non-SVE targets,
since we wouldn't be able to undefine VG for non-SVE targets.
As a compromise, it seems we can safely fetch the VG register value from
ptrace. The value in the kernel is likely the updated value anyway.
This patch fixed all the failures i saw in the testsuite and caused no further
regressions.
gdb/ChangeLog:
2020-03-19 Luis Machado <luis.machado@linaro.org>
* nat/aarch64-sve-linux-ptrace.c (aarch64_sve_set_vq): If vg is not
valid, fetch vg value from ptrace.
Add gdb_ptrace() that wraps the ptrace(2) API and correctly passes
the pid,lwp pair to the calls on NetBSD; and the result of
get_ptrace_pid() on other BSD Operating Systems.
gdb/ChangeLog:
* x86-bsd-nat.c (gdb_ptrace): New.
* (x86bsd_dr_set): Add new argument `ptid'.
* (x86bsd_dr_get, x86bsd_dr_set, x86bsd_dr_set_control,
x86bsd_dr_set_addr): Update.
process_symbol_table () has
unsigned long num_syms;
...
for (si = 0, psym = symtab; si < num_syms; si++, psym++)
We should use unsigned long to iterate over num_syms.
* readelf.c (process_symbol_table): Use unsigned long for si.
In this commit:
commit 24ed6739b6
Date: Thu Jan 30 14:35:40 2020 +0000
gdb/remote: Restore support for 'S' stop reply packet
A regression was introduced such that the W and X packets would give a
warning in some cases. The warning was:
warning: multi-threaded target stopped without sending a thread-id, using first non-exited thread
This problem would arise when:
1. The multi-process extensions to the remote protocol were not
being used, and
2. The inferior has multiple threads.
In this case when the W (or X) packet arrives the ptid of the
stop_reply is set to null_ptid, then when we arrive in
process_stop_reply GDB spots that we have multiple non-exited theads,
but the stop event didn't specify a thread-id.
The problem with this is that the W (and X) packets are actually
process wide events, they apply to all threads. So not specifying a
thread-id is not a problem, in fact, the best these packets allow is
for the remote to specify a process-id, not a thread-id.
If we look at how the W (and X) packets deal with a specified
process-id, then what happens is GDB sets to stop_reply ptid to a
value which indicates all threads in the process, this is done by
creating a value `ptid_t (pid)`, which sets the pid field of the
ptid_t, but leaves the tid field as 0, indicating all threads.
So, this commit does the same thing for the case where there is not
process-id specified. In process_stop_reply we not distinguish
between stop events that apply to all threads, and those that apply to
only one. If the stop event applies to only one thread then we treat
it as before. If, however, the stop event applies to all threads,
then we find the first non-exited thread, and use the pid from this
thread to create a `ptid_t (pid)` value.
If the target has multiple inferiors, and receives a process wide
event without specifying a process-id GDB now gives this warning:
warning: multi-inferior target stopped without sending a process-id, using first non-exited inferior
gdb/ChangeLog:
* remote.c (remote_target::process_stop_reply): Handle events for
all threads differently.
gdb/testsuite/ChangeLog:
* gdb.server/exit-multiple-threads.c: New file.
* gdb.server/exit-multiple-threads.exp: New file.
This commit adds a test that builds a mixed language stack, the stack
contains frames of Fortran, C, and C++. The test prints the backtrace
and explores the stack printing arguments of different types in frames
of different languages.
The core of the test is repeated with GDB's language set to auto,
fortran, c, and c++ in turn to ensure that GDB is happy to print
frames and frame arguments when the language is set to a value that
doesn't match the frame language.
This test currently passes, and there are no known bugs in this area.
The aim of this commit is simply to increase test coverage, as I don't
believe this functionality is currently tested.
gdb/testsuite/ChangeLog:
* gdb.fortran/mixed-lang-stack.c: New file.
* gdb.fortran/mixed-lang-stack.cpp: New file.
* gdb.fortran/mixed-lang-stack.exp: New file.
* gdb.fortran/mixed-lang-stack.f90: New file.
Consider debugging the following C++ program:
struct object
{ int a; };
typedef object *object_p;
static int
get_value (object_p obj)
{
return obj->a;
}
int
main ()
{
object obj;
obj.a = 0;
return get_value (&obj);
}
Now in a GDB session:
(gdb) complete break get_value
break get_value(object*)
break get_value(object_p)
Or:
(gdb) break get_va<TAB>
(gdb) break get_value(object<RETURN>
Function "get_value(object" not defined.
Make breakpoint pending on future shared library load? (y or [n]) n
The reason this happens is that we add completions based on the
msymbol names and on the symbol names. For C++ both of these names
include the parameter list, however, the msymbol names have some
differences from the symbol names, for example:
+ typedefs are resolved,
+ whitespace rules are different around pointers,
+ the 'const' keyword is placed differently.
What this means is that the msymbol names and symbol names appear to
be completely different to GDB's completion tracker, and therefore to
readline when it offers the completions.
This commit builds on the previous commit which reworked the
completion_tracker class. It is now trivial to add a
remove_completion member function, this is then used along with
cp_canonicalize_string_no_typedefs to remove the msymbol aliases from
the completion tracker as we add the symbol names.
Now, for the above program GDB only presents a single completion for
'get_value', which is 'get_value(object_p)'.
It is still possible to reference the symbol using the msymbol name,
so a user can manually type out 'break get_value (object *)' if they
wish and will get the expected behaviour.
I did consider adding an option to make this alias exclusion optional,
in the end I didn't bother as I didn't think it would be very useful,
but I can easily add such an option if people think it would be
useful.
gdb/ChangeLog:
* completer.c (completion_tracker::remove_completion): Define new
function.
* completer.h (completion_tracker::remove_completion): Declare new
function.
* symtab.c (completion_list_add_symbol): Remove aliasing msymbols
when adding a C++ function symbol.
gdb/testsuite/ChangeLog:
* gdb.linespec/cp-completion-aliases.cc: New file.
* gdb.linespec/cp-completion-aliases.exp: New file.
Change-Id: Ie5c7c9fc8ecf973072cfb4a9650867104bf7f50c
In this commit I rewrite how the completion tracker tracks the
completions, and builds its lowest common denominator (LCD) string.
The LCD string is now built lazily when required, and we only track
the completions in one place, the hash table, rather than maintaining
a separate vector of completions.
The motivation for these changes is that the next commit will add the
ability to remove completions from the list, removing a completion
will invalidate the LCD string, so we need to keep hold of enough
information to recompute the LCD string as needed.
Additionally, keeping the completions in a vector makes removing a
completion expensive, so better to only keep the completions in the
hash table.
This commit doesn't add any new functionality itself, and there should
be no user visible changes after this commit.
For testing, I ran the testsuite as usual, but I also ran some manual
completion tests under valgrind, and didn't get any reports about
leaked memory.
gdb/ChangeLog:
* completer.c (completion_tracker::completion_hash_entry): Define
new class.
(advance_to_filename_complete_word_point): Call
recompute_lowest_common_denominator.
(completion_tracker::completion_tracker): Call discard_completions
to setup the hash table.
(completion_tracker::discard_completions): Allow for being called
from the constructor, pass new equal function, and element deleter
when constructing the hash table. Initialise new class member
variables.
(completion_tracker::maybe_add_completion): Remove use of
m_entries_vec, and store more information into m_entries_hash.
(completion_tracker::recompute_lcd_visitor): New function, most
content taken from...
(completion_tracker::recompute_lowest_common_denominator):
...here, this now just visits each item in the hash calling the
above visitor.
(completion_tracker::build_completion_result): Remove use of
m_entries_vec, call recompute_lowest_common_denominator.
* completer.h (completion_tracker::have_completions): Remove use
of m_entries_vec.
(completion_tracker::completion_hash_entry): Declare new class.
(completion_tracker::recompute_lowest_common_denominator): Change
function signature.
(completion_tracker::recompute_lcd_visitor): Declare new function.
(completion_tracker::m_entries_vec): Delete.
(completion_tracker::m_entries_hash): Initialize to NULL.
(completion_tracker::m_lowest_common_denominator_valid): New
member variable.
(completion_tracker::m_lowest_common_denominator_max_length): New
member variable.
Change-Id: I9d1db52c489ca0041b8959ca0d53b7d3af8aea72
When running test-case gdb.opt/inline-locals.exp, I get:
...
Running src/gdb/testsuite/gdb.opt/inline-locals.exp ...
KPASS: gdb.opt/inline-locals.exp: info locals above bar 2 (PRMS gdb/xyz)
KPASS: gdb.opt/inline-locals.exp: info locals above bar 3 (PRMS gdb/xyz)
...
I've opened PR25695 - 'abstract and concrete variable listed both with "info
locals"' to refer to in the PRMS field, and this patch adds that reference.
Furthermore, I noticed that while I see KPASSes, given the problem description
the tests should actually be KFAILs. This patch also fixes that.
Tested on x86_64-linux. With gcc 7.5.0, I get 2 KFAILs. With clang 5.0.2,
the tests pass.
gdb/testsuite/ChangeLog:
2020-03-19 Tom de Vries <tdevries@suse.de>
* gdb.opt/inline-locals.exp: Add kfail PR number. Make kfail matching
more precise.
Better than warning about bfd types, just don't include bfd.h and
warn against including the header again.
* elfcomm.c: Don't include bfd.h or bucomm.h.
(program_name): Declare.
(process_archive_index_and_symbols): Replace bfd_boolean with int,
and substitute FALSE and TRUE.
(setup_archive, setup_nested_archive): Likewise.
* elfcomm.h: Likewise.
Add a test-case that tests whether we can set a breakpoint on an inlined
inline function in CU for which the partial symtab has not yet been expanded.
Tested on x86_64-linux, with gcc 4.8.5, gcc-7.5.0, gcc-10.0.1, and clang
5.0.2.
gdb/testsuite/ChangeLog:
2020-03-18 Tom de Vries <tdevries@suse.de>
* gdb.dwarf2/break-inline-psymtab-2.c: New test.
* gdb.dwarf2/break-inline-psymtab.c: New test.
* gdb.dwarf2/break-inline-psymtab.exp: New file.
