Simon Marchi [Tue, 5 Dec 2017 21:30:28 +0000 (16:30 -0500)]
Split tdesc_type into multiple classes
This patch makes tdesc_type an abstract base class and creates three
subclasses:
- tdesc_type_builtin, for builtin types
- tdesc_type_vector, for vector types
- tdesc_type_with_fields, for struct, union, flag and enum types
This allows getting rid of the union in tdesc_type and to not allow the
std::vector separately. I tried to go further and create separate
classes for struct, union, flag and enum, but it proved too difficult.
One problem is that from the point of the of the target description
code, the types tdesc_type_* are opaque (only forward-declared).
Therefore, it doesn't know about inheritance relationship between those
classes. This makes it impossible to make functions that accept a
pointer to a base class and pass a pointer to a derived class, for
example. I think this patch here is a good compromise, and if somebody
wants to improve things further, the door is open.
A make_gdb_type virtual pure method is added to tdesc_type, which
replaces the current tdesc_gdb_type function. Calling this method on a
tdesc_type returns the corresponding built gdb type.
Simon Marchi [Tue, 5 Dec 2017 21:30:27 +0000 (16:30 -0500)]
Make tdesc_arch_data::arch_regs an std::vector
Make tdesc_arch_data::arch_regs be an std::vector of tdesc_arch_reg
objects.
On particularity is that the tdesc_arch_data linked to a gdbarch is
allocated on the gdbarch's obstack. To be safe, I did not change it and
called placement-new on the area returned by OBSTACK_ZALLOC.
gdb/ChangeLog:
* target-descriptions.c (tdesc_arch_reg): Remove typedef.
(struct tdesc_arch_reg): Add constructor.
(DEF_VEC_O (tdesc_arch_reg)): Remove.
(struct tdesc_arch_data): Initialize fields.
<arch_regs>: Change type to std::vector.
(target_find_description): Adjust.
(tdesc_find_type): Adjust.
(tdesc_data_init): Call tdesc_arch_data constructor.
(tdesc_data_alloc): Allocate tdesc_arch_data with new.
(tdesc_data_cleanup): Free data with delete.
(tdesc_numbered_register): Adjust.
(tdesc_find_arch_register): Adjust.
(tdesc_use_registers): Adjust.
Simon Marchi [Tue, 5 Dec 2017 21:30:26 +0000 (16:30 -0500)]
Make tdesc_type::u::u::fields an std::vector
This patch makes the tdesc_type::u::u::fields an std::vector of
tdesc_type_field. The difficulty here is that the vector is part of a
union. Because of this, I made fields a pointer to a vector, and
instantiate/destroy the vector if the type is one that uses this member
of the union
The field tdesc_type_field::name is changed to an std::string at the
same time.
gdb/ChangeLog:
* target-descriptions.c (tdesc_type_field): Remove typedef.
(DEF_VEC_O (tdesc_type_field)): Remove.
(struct tdesc_type_field): Add constructor.
<name>: Change type to std::string.
(struct tdesc_type) <tdesc_type>: Instantiate vector if the type
kind uses it.
<~tdesc_type>: Destroy vector if the type kind uses it.
<u::u::fields>: Change type to std::vector.
(tdesc_gdb_type): Adjust.
(tdesc_add_field): Adjust.
(tdesc_add_typed_bitfield): Adjust.
(tdesc_add_field): Adjust.
(tdesc_add_enum_value): Adjust.
(class print_c_tdesc) <visit>: Adjust.
Simon Marchi [Tue, 5 Dec 2017 21:30:26 +0000 (16:30 -0500)]
Make tdesc_type::name an std::string
This patch makes tdesc_type::name an std::string. This way, we don't
need to free it manually in ~tdesc_type. I think the comment on top of
name is not correct, the string is always malloc'ed.
Simon Marchi [Tue, 5 Dec 2017 21:30:25 +0000 (16:30 -0500)]
Make tdesc_feature::types an std::vector
This patch makes tdesc_feature::types an std::vector of unique_ptr of
tdesc_type. This way, we don't need to manually free the objects and
the vector in ~tdesc_feature.
Simon Marchi [Tue, 5 Dec 2017 21:30:25 +0000 (16:30 -0500)]
Make tdesc_reg string fields std::string
Make the name, group and type fields of tdesc_reg std::strings. This
way, we don't have to manually free them in ~tdesc_reg.
Doing so results in a small change in the generated tdesc. Instead of
passing an empty string for the group parameter of tdesc_create_reg, the
two modified tdesc now pass NULL. The end result should be the same.
gdb/ChangeLog:
* target-descriptions.c (struct tdesc_reg) <tdesc_reg>: Change
type of name_ parameter, adjust to std::string change.
<name, group, type>: Change type to std::string.
<~tdesc_reg>: Replace with default implementation.
<operator==>: Adjust.
(tdesc_find_register_early): Adjust.
(tdesc_register_name): Adjust.
(tdesc_register_type): Adjust.
(tdesc_register_in_reggroup_p): Adjust.
(class print_c_tdesc) <visit>: Adjust.
(class print_c_feature) <visit>: Adjust.
Simon Marchi [Tue, 5 Dec 2017 21:30:24 +0000 (16:30 -0500)]
Make tdesc_feature::registers an std::vector
This patch makes tdesc_feature::registers an std::vector of unique_ptr
to tdesc_reg. This way, we don't have to manually free the tdesc_reg
objects and the vector in the tdesc_feature destructor.
Simon Marchi [Tue, 5 Dec 2017 21:30:23 +0000 (16:30 -0500)]
Make target_desc::features an std::vector
This patch makes target_desc to be a vector of unique_ptr to
tdesc_feature objects. This way, we don't have to manually free the
features and the vector in the target_desc destructor.
Simon Marchi [Tue, 5 Dec 2017 21:30:22 +0000 (16:30 -0500)]
Make target_desc::compatible an std::vector
This patch changes target_desc::compatible to be a vector of
bfd_arch_info *. This way, we don't need to manually free the vector in
the target_desc destructor.
Simon Marchi [Tue, 5 Dec 2017 21:30:22 +0000 (16:30 -0500)]
Make target_desc::properties an std::vector
This patch changes target_desc::properties to be a vector of property
objects. This way, we don't need to manually free the property members
as well as the property objects themselves.
Simon Marchi [Tue, 5 Dec 2017 21:15:08 +0000 (16:15 -0500)]
Redefine gdb_static_assert as static_assert
Since we use C++11, we can use static_assert instead doing the trick
that makes a negative-sized array if the expression is false.
static_assert is built in the language and gives clearer error messages.
To avoid modifying the usages of gdb_static_assert, redefine
gdb_static_assert in terms of static_assert, passing an empty message.
If we want to add an assert with a message, it's always possible to use
static_assert directly.
gdb/ChangeLog:
* common/gdb_assert.h (gdb_static_assert): Redefine using
static_assert.
Sergio Lopez [Mon, 4 Dec 2017 08:17:14 +0000 (09:17 +0100)]
Implement "-a" command line option for gcore
With the new "-a" command line option, the user may request gcore to
actually dump all present memory mappings. The actual effect of this
argument is OS dependent.
On GNU/Linux, it will disable use-coredump-filter and enable
dump-excluded-mappings.
Sergio Lopez [Mon, 4 Dec 2017 08:17:12 +0000 (09:17 +0100)]
Implement 'set dump-excluded-mappings' command
Commit df8411da087dc05481926f4c4a82deabc5bc3859 implemented support for
checking /proc/PID/coredump_filter, and also changed gcore behavior to
unconditionally honor the VM_DONTDUMP flag, preventing sections marked
as such for being dumped into the core file.
This patch implements the 'set dump-excluded-mappings' command for
instructing gdb to ignore the VM_DONTDUMP flag. Combined with 'set
use-coredump-filter', this allows the user to restore the old behavior,
dumping all sections (except the ones marked as IO) unconditionally.
* linux-tdep.c (dump_excluded_mappings): New variable.
(dump_mapping_p): Use dump_excluded_mappings variable.
(_initialize_linux_tdep): New command 'set dump_excluded_mappings'.
PR gdb/22499 is about a latent bug exposed by the switch to "maint set
target-non-stop on" by default on x86-64 GNU/Linux, a while ago. With
that on, GDB is also preferring to use displaced-stepping by default.
The testcase in the bug is failing because GDB ends up incorrectly
displaced-stepping over a RIP-relative VEX-encoded instruction, like
this:
While RIP-relative instructions need adjustment when relocated to the
scratch pad, GDB ends up just copying VEX-encoded instructions to the
scratch pad unmodified, with the end result that the inferior ends up
executing an instruction that fetches/writes memory from the wrong
address...
This patch teaches GDB about the VEX-encoding prefixes, fixing the
problem, and adds a testcase that fails without the GDB fix.
I think we may need a similar treatment for EVEX-encoded instructions,
but I didn't address that simply because I couldn't find any
EVEX-encoded RIP-relative instruction in the gas testsuite. In any
case, this commit is forward progress as-is already.
gdb/ChangeLog:
2017-12-04 Pedro Alves <palves@redhat.com>
PR gdb/22499
* amd64-tdep.c (amd64_insn::rex_offset): Rename to...
(amd64_insn::enc_prefix_offset): ... this, and tweak comment.
(vex2_prefix_p, vex3_prefix_p): New functions.
(amd64_get_insn_details): Adjust to rename. Also skip VEX2 and
VEX3 prefixes.
(fixup_riprel): Set VEX3.!B.
gdb/testsuite/ChangeLog:
2017-12-04 Pedro Alves <palves@redhat.com>
PR gdb/22499
* gdb.arch/amd64-disp-step-avx.S: New file.
* gdb.arch/amd64-disp-step-avx.exp: New file.
Pedro Alves [Sun, 3 Dec 2017 17:50:43 +0000 (12:50 -0500)]
Make make-target-delegates grok namespace scope op and template params
The next patch will want to use gdb::array_view<int> as parameter type
of a target_ops method. However, that runs into a
make-target-delegates limitation: target_debug_foo calls in
target-delegates.c for parameters/return types with namespace scope
operators ("::") or template parameters, end up looking like:
@@ -1313,9 +1313,7 @@ debug_set_syscall_catchpoint (struct target_ops *self, int arg1, int arg2, int a
fputs_unfiltered (", ", gdb_stdlog);
target_debug_print_int (arg3);
fputs_unfiltered (", ", gdb_stdlog);
- target_debug_print_int (arg4);
- fputs_unfiltered (", ", gdb_stdlog);
- target_debug_print_int_p (arg5);
+ target_debug_print_gdb::array_view<const_int> (arg4);
which obviously isn't something that compiles. The problem is that
make-target-delegates wasn't ever taught that '::', '<', and '>' can
appear in parameter/return types. You could work around it by hidding
the unsupported characters behind a typedef in the target method
declaration, or by using an explicit TARGET_DEBUG_PRINTER, but it's
better to just remove the limitation.
While at it, also fix an "abuse" of reserved identifiers.
gdb/ChangeLog:
* make-target-delegates (munge_type): Also munge '<', '>', and
':'. Avoid double underscores in identifiers, and trailing
underscores.
* target-debug.h
(target_debug_print_VEC_static_tracepoint_marker_p__p): Rename to
...
(target_debug_print_VEC_static_tracepoint_marker_p_p): ... this.
* target-delegates.c: Regenerate.
Pedro Alves [Sun, 3 Dec 2017 15:32:08 +0000 (15:32 +0000)]
Fix gdb.threads/process-dies-while-detaching.exp
I noticed [1] a test bug in gdb.threads/process-dies-while-detaching.exp.
Simplified, the test code in question looks somewhat like this:
~~~
# Detach from a process, and ensure that it exits after detaching.
# This relies on inferior I/O.
proc detach_and_expect_exit {test} {
gdb_test_multiple "detach" $test ....
set saw_prompt 0
set saw_inf_exit 0
while { !$saw_prompt && !$saw_inf_exit } {
gdb_test_multiple "" $test {
-re "exited, status=0" {
set saw_inf_exit 1
}
-re "$gdb_prompt " {
set saw_prompt 1
}
}
}
pass $test
}
~~~
The bug is in the while loop's condition. We want to make sure we see
both the inferior output and the prompt, so the loop's test should be:
- while { !$saw_prompt && !$saw_inf_exit } {
+ while { !$saw_prompt || !$saw_inf_exit } {
If we just fix that, the test starts failing though, because it
exposes a couple latent problems:
- When called from test_detach_killed_outside, the parent doesn't
print "exited, status=0", because in that case the child dies with a
signal, and so detach_and_expect_exit times out.
Fix it by making the parent print "signaled, sig=9" in that case,
and have the .exp expect it.
- When testing against --target_board=native-gdbserver, sometimes we'd
get this:
ERROR: Process no longer exists
ERROR: : spawn id exp9 not open
while executing
"expect {
-i exp8 -timeout 220
-i $server_spawn_id
eof {
pass $test
wait -i $server_spawn_id
unset server_spawn_id
}
timeout {
..."
("uplevel" body line 1)
invoked from within
"uplevel $body" NONE : spawn id exp9 not open
The problem is that:
- inferior_spawn_id and server_spawn_id are the same when testing
with gdbserver.
- gdbserver exits after "detach", so we get an eof for
$inferior_spawn_id in the loop in detach_and_expect_exit.
That's the first "ERROR: Process no longer exists".
- and then when we reach test_server_exit, server_spawn_id
is already closed (because server_spawn_id==inferior_spawn_id).
To handle this, make the loop in detach_and_expect_exit use an
indirect spawn id list and remove $inferior_spawn_id from the list
as soon as we got the inferior output we're expecting, so that the
"eof" is left unprocessed until we reach test_server_exit.
[1] I changed GDB in a way that should have made the test fail, but it
didn't.
gdb/testsuite/ChangeLog:
2017-12-03 Pedro Alves <palves@redhat.com>
* gdb.threads/process-dies-while-detaching.c: Include <errno.h>
and <string.h>.
(parent_function): Print distinct messages when waitpid fails, or
the child exits with a signal, or the child exits for an unhandled
reason.
* gdb.threads/process-dies-while-detaching.exp
(detach_and_expect_exit): New 'inf_output_re' parameter and use
it. Wait for both inferior output and GDB's prompt. Use an
indirect spawn id list.
(do_detach): New parameter 'child_exit'. Use it to compute
expected inferior output.
(test_detach, test_detach_watch, test_detach_killed_outside):
Adjust to pass down the expected child exit kind.
Simon Marchi [Sun, 3 Dec 2017 01:36:39 +0000 (20:36 -0500)]
Remove usages of find_inferior calling not_stopped_callback
Replace with find_thread. Writing a lambda inline in directly in the if
conditions would be a bit messy, so I chose to assign them to variables
instead.
gdb/gdbserver/ChangeLog:
* linux-low.c (not_stopped_callback): Return bool, take filter
argument directly.
(linux_wait_for_event_filtered): Use find_thread.
(linux_wait_1): Likewise.
Simon Marchi [Sun, 3 Dec 2017 01:36:38 +0000 (20:36 -0500)]
Remove usage of find_inferior in find_lwp_pid
Replace with find_thread. We could almost use find_thread_ptid, except
that find_lwp_pid uses the pid of the input ptid of the lwp is 0, so the
behavior is not quite the same.
gdb/gdbserver/ChangeLog:
* linux-low.c (same_lwp): Remove.
(find_lwp_pid): Use find_thread.
Simon Marchi [Sun, 3 Dec 2017 01:36:34 +0000 (20:36 -0500)]
Remove find_inferior_id
Remove find_inferior_id, replacing its usages with find_thread_ptid.
find_thread_ptid was implemented using find_inferior_id, so move the
implementation there instead.
Simon Marchi [Sun, 3 Dec 2017 01:36:32 +0000 (20:36 -0500)]
Remove usage of find_inferior in regcache_invalidate_pid
Replace with for_each_thread with pid filtering.
regcache_invalidate_one is not longer needed, as it was only used to
filter the pid. We can call regcache_invalidate_thread directly.
gdb/gdbserver/ChangeLog:
* regcache.c (regcache_invalidate_one): Remove.
(regcache_invalidate_pid): use for_each_thread.
Cary Coutant [Wed, 15 Feb 2017 08:25:29 +0000 (00:25 -0800)]
Disallow --incremental with -pie and force -no-pie for incremental tests.
This is a partial fix for the gold testsuite failures documented in
PR 21090. The use of -fpie triggers some mov-to-lea optimizations that
are not compatible with incremental linking, so those optimizations need
to be disabled. We also diagnose the attempt to use -pie with incremental
linking, and force -no-pie for the incremental tests in case the build has
been configured to have GCC pass -pie all the time.
We still have a problem where compiling with -fpie results in some GOT
entries even when linking with -no-pie. This combination still causes test
failures because we are not updating the GOT entries in an incremental update
link.
Cary Coutant [Sat, 2 Dec 2017 06:03:09 +0000 (22:03 -0800)]
Fix incremental linking failure with GCC 7+.
With the new compiler, we're running out of patch space for the .eh_frame
section. To workaround that issue, we compile the before and after versions
both with no unwind tables.
gold/
PR gold/22309
* testsuite/Makefile.am (two_file_test_1_v1_ndebug.o): Compile with
no EH information.
(two_file_test_1_ndebug.o): Likewise.
* testsuite/Makefile.in: Regenerate.
* testsuite/two_file_test_1.cc: Touch to force recompilation with new
flags.
* testsuite/two_file_test_1_v1.cc: Likewise.
The purpose of this concept is to turn the load of debugging
information off, either globally (via the '--readnever' option), or
objfile-specific. The implementation proposed here is an extension of
the patch distributed with Fedora GDB; looking at the Fedora patch
itself and the history, one can see some reasons why it was never
resubmitted:
- The patch appears to have been introduced as a workaround, at
least initially;
- The patch is far from perfect, as it simply shunts the load of
DWARF debugging information, without really worrying about the
other debug format.
- Who really does non-symbolic debugging anyways?
One use of this feature is when a user simply wants to do the
following sequence: attach, dump core, detach. Loading the debugging
information in this case is an unnecessary cause of delay.
This patch expands the version shipped with Fedora GDB in order to
make the feature available for all the debuginfo backends, not only
for DWARF. It also implements a per-objfile flag which can be
activated by using the "-readnever" command when using the
'add-symbol-file' or 'symbol-file' commands.
It's also worth mentioning that this patch tests whether GDB correctly
fails to initialize if both '--readnow' and '--readnever' options are
passed.
Tested on the BuildBot.
gdb/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* NEWS (Changes since GDB 8.0: Mention new '--readnever'
feature.
* coffread.c (coff_symfile_read): Do not map over sections with
'coff_locate_sections' if readnever is on.
* dwarf2read.c (dwarf2_has_info): Return 0 if
readnever is on.
* elfread.c (elf_symfile_read): Do not map over sections with
'elf_locate_sections' if readnever is on.
* main.c (validate_readnow_readnever): New function.
(captured_main_1): Add support for --readnever.
(print_gdb_help): Document --readnever.
* objfile-flags.h (enum objfile_flag) <OBJF_READNEVER>: New
flag.
* symfile.c (readnever_symbol_files): New global.
(symbol_file_add_with_addrs): Set 'OBJF_READNEVER' when
'READNEVER_SYMBOL_FILES' is set.
(validate_readnow_readnever): New function.
(symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(add_symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(_initialize_symfile): Document new '-readnever' option for
both 'symbol-file' and 'add-symbol-file' commands.
* top.h (readnever_symbol_files): New extern global.
* xcoffread.c (xcoff_initial_scan): Do not read debug
information if readnever is on.
gdb/doc/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.texinfo (File Options): Document --readnever.
(Commands to Specify Files): Likewise, for 'symbol-file' and
'add-symbol-file'.
gdb/testsuite/ChangeLog:
2017-12-01 Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
Pedro Alves <palves@redhat.com>
* gdb.base/readnever.c, gdb.base/readnever.exp: New files.
Make '{add-,}symbol-file' not care about the position of command line arguments
This is a bug that's been detected while doing the readnever work.
If you use 'symbol-file' or 'add-symbol-file', the position of each
argument passed to the command matters. This means that if you do:
(gdb) symbol-file -readnow /foo/bar
The symbol file specified will (correctly) have all of its symbols
read by GDB (because of the -readnow flag). However, if you do:
(gdb) symbol-file /foo/bar -readnow
GDB will silently ignore the -readnow flag, because it was specified
after the filename. This is not a good thing to do and may confuse
the user.
To address that, I've modified the argument parsing mechanisms of
symbol_file_command and add_symbol_file_command to be
"position-independent". I have also added one error call at the end
of add_symbol_file_command's argument parsing logic, which now clearly
complains if no filename has been specified. Both commands now
support the "--" option to stop argument processing.
This patch provides a testcase for both commands, in order to make
sure that the argument order does not matter. It has been
regression-tested on BuildBot.
* symfile.c (symbol_file_command): Call
'symbol_file_add_main_1' only after processing all command
line options.
(add_symbol_file_command): Modify logic to make arguments
position-independent.
Joel Brobecker [Fri, 1 Dec 2017 20:45:30 +0000 (15:45 -0500)]
(Ada) GDB crash printing expression with type casting
One of our users reported that trying to print the following expression,
caused GDB to SEGV:
(gdb) print some_package.some_type (val)
In this particular instance, the crash occurred inside ada_args_match
because it is given a NULL "func", leading to the SEGV because of:
struct type *func_type = SYMBOL_TYPE (func);
This NULL symbol comes from a list of symbols which was given to
ada_resolve_function (parameter called "syms") which then iterates
over each of them to discard the ones that don't match the actuals:
for (k = 0; k < nsyms; k += 1)
{
struct type *type = ada_check_typedef (SYMBOL_TYPE (syms[k].symbol));
What's really interesting is that, when entering the block above for
the first time, all entries in SYMS have a valid (non-NULL) symbol.
However, once we return from the call to ada_check_typedef, the first
entry of our SYMS table gets set to all zeros:
(gdb) p syms[0]
$2 = {symbol = 0x0, block = 0x0}
Hence the call to ada_args_match with a NULL symbol, and the ensuing
SEGV.
To find out why this happen, we need to step back a little and look
at how syms was allocated. This list of symbols comes from a symbol
lookup, which means ada_lookup_symbol_list_worker. We have our first
hint when we look at the function's documentation and see:
This vector is transient---good only to the next call of
ada_lookup_symbol_list.
Implementation-wise, this is done by using a static global obstack,
which we just re-initialize each time ada_lookup_symbol_list_worker
gets called:
This property was probably established in order to facilitate the use
of the returned vector, since the users of that function would not have
to worry about releasing that memory when no longer needed. However,
I found during this investigation that it is all to easy to indirectly
trigger another symbol lookup while still using the results of a previous
lookup.
In our particular case, there is the call to ada_check_typedef, which
leads to check_typedef. As it happens, my first symbol had a type which
was a typedef to a stub type, so check_typedef calls lookup_symbol to
find the non-stub version. This in turn eventually leads us back to
ada_lookup_symbol_list_worker, where the first thing it does is free
the memory area when our list of symbols have been residing and then
recreates a new one. in other words, SYMS then becomes a dangling
pointer!
This patch fixes the issue by having ada_lookup_symbol_list_worker
return a copy of the list of symbols, with the responsibility of
deallocating that list now transfered to the users of that list.
More generally speaking, it is absolutely amazing that we haven't seen
consequences of this issue before. This can happen fairly frequently.
For instance, I found that ada-exp.y::write_var_or_type calls
ada_lookup_symbol_list, and then, while processing that list, calls
select_possible_type_sym, which leads to ada_prefer_type, eventually
leading to ada_check_typedef again (via eg. ada_is_array_descriptor_type).
Even more amazing is the fact that, while I was able to produce multiple
scenarios where the corruption occurs, none of them leads to incorrect
behavior at the user level. In other words, it requires a very precise
set of conditions for the corruption to become user-visible, and
despite having a megalarge program where the crash occured, using that
as a template for creating a reproducer did not work (pb goes away).
This is why this patch does not come with a reproducer. On the other hand,
this should not be a problem in terms of testing coverage, as the changes
are made in common areas which, at least for the most part, are routinely
exercised during testing.
gdb/ChangeLog:
* ada-lang.c (symbol_list_obstack): Delete.
(resolve_subexp): Make sure "candidates" gets xfree'ed.
(ada_lookup_symbol_list_worker): Remove the limitation that
the result is only good until the next call, now making it
the responsibility of the caller to free the result when no
longer needed. Adjust the function's intro comment accordingly.
(ada_lookup_symbol_list): Adjust the function's intro comment.
(ada_iterate_over_symbols): Make sure "results" gets xfree'ed.
(ada_lookup_encoded_symbol, get_var_value): Likewise.
(_initialize_ada_language): Remove symbol_list_obstack
initialization.
* ada-exp.y (block_lookup): Make sure "syms" gets xfree'ed.
(write_var_or_type, write_name_assoc): Likewise.
Fix internal error from command line with unbalanced --start-lib/--end-lib.
The problem is that while the command line isn't trivially empty,
it contains no input files. As gold tries to configure the number
of threads to use based on the number of input files, this causes
the assertion failure above. Fix this problem by making the logic
in gold.cc more robust and also adding a better error message
about --start-lib to options.cc.
gold/
PR gold/22406
* gold.cc (queue_initial_tasks) Check for number of real input files.
* options.cc (Command_line::process) Check for unterminated --start-lib
options.
* testsuite/Makefile.am: Add new test script.
* testsuite/Makefile.in: Regenerate.
* testsuite/check_empty_command_lines.sh: New test script.
Stephen Crane [Fri, 1 Dec 2017 20:10:02 +0000 (12:10 -0800)]
Fix problem where undef can fail to trigger archive rescan.
If a shared library contains an undefined symbol and LTO adds
a new reference to that same undefined symbol, the reference in the new
object added by the plugin would not trigger a rescan of the archive
containing the symbol.
2017-11-17 Stephen Crane <sjc@immunant.com>
gold/
PR gold/22448
* symtab.cc (Symbol_table::add_from_object): Only rescan for
undefined symbols in regular, not dynamic, objects.
The purpose of this concept is to turn the load of debugging
information off, either globally (via the '--readnever' option), or
objfile-specific. The implementation proposed here is an extension of
the patch distributed with Fedora GDB; looking at the Fedora patch
itself and the history, one can see some reasons why it was never
resubmitted:
- The patch appears to have been introduced as a workaround, at
least initially;
- The patch is far from perfect, as it simply shunts the load of
DWARF debugging information, without really worrying about the
other debug format.
- Who really does non-symbolic debugging anyways?
One use of this feature is when a user simply wants to do the
following sequence: attach, dump core, detach. Loading the debugging
information in this case is an unnecessary cause of delay.
This patch expands the version shipped with Fedora GDB in order to
make the feature available for all the debuginfo backends, not only
for DWARF. It also implements a per-objfile flag which can be
activated by using the "-readnever" command when using the
'add-symbol-file' or 'symbol-file' commands.
It's also worth mentioning that this patch tests whether GDB correctly
fails to initialize if both '--readnow' and '--readnever' options are
passed.
Tested on the BuildBot.
gdb/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* NEWS (Changes since GDB 8.0: Mention new '--readnever'
feature.
* coffread.c (coff_symfile_read): Do not map over sections with
'coff_locate_sections' if readnever is on.
* dwarf2read.c (dwarf2_has_info): Return 0 if
readnever is on.
* elfread.c (elf_symfile_read): Do not map over sections with
'elf_locate_sections' if readnever is on.
* main.c (validate_readnow_readnever): New function.
(captured_main_1): Add support for --readnever.
(print_gdb_help): Document --readnever.
* objfile-flags.h (enum objfile_flag) <OBJF_READNEVER>: New
flag.
* symfile.c (readnever_symbol_files): New global.
(symbol_file_add_with_addrs): Set 'OBJF_READNEVER' when
'READNEVER_SYMBOL_FILES' is set.
(validate_readnow_readnever): New function.
(symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(add_symbol_file_command): Handle '-readnever' option.
Call 'validate_readnow_readnever'.
(_initialize_symfile): Document new '-readnever' option for
both 'symbol-file' and 'add-symbol-file' commands.
* top.h (readnever_symbol_files): New extern global.
* xcoffread.c (xcoff_initial_scan): Do not read debug
information if readnever is on.
gdb/doc/ChangeLog:
2017-12-01 Andrew Cagney <cagney@redhat.com>
Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.texinfo (File Options): Document --readnever.
(Commands to Specify Files): Likewise, for 'symbol-file' and
'add-symbol-file'.
gdb/testsuite/ChangeLog:
2017-12-01 Joel Brobecker <brobecker@adacore.com>
Sergio Durigan Junior <sergiodj@redhat.com>
Pedro Alves <palves@redhat.com>
* gdb.base/readnever.c, gdb.base/readnever.exp: New files.
Tom Tromey [Thu, 30 Nov 2017 18:49:27 +0000 (11:49 -0700)]
Fix dependency tracking for objects in subdirectories
On irc, Pedro pointed out that dependencies for objects in
subdirectories didn't seem to be working.
The bug was that the "-include" for .deps files was using the wrong file
name for subdirectory objects; e.g., for cli/cli-decode.o it was trying
to open .deps/cli/cli-decode.o, whereas the correct file is
cli/.deps/cli-decode.o.
This patch changes how the dep files are found. Tested by touching a
source file and rebuilding cli/cli-decode.o.
2017-12-01 Tom Tromey <tom@tromey.com>
* Makefile.in (all_deps_files): New variable.
Include .Po files using all_deps_files.
Yao Qi [Fri, 1 Dec 2017 11:34:14 +0000 (11:34 +0000)]
Replace mail address with the URL in copyright header
The copyright header in most of GDB files were changed from mail address
to the URL in the conversion to GPLv3 in Aug 2007. However, some files
still use mail address instead of the URL. This patch fixes them.
Alan Modra [Fri, 1 Dec 2017 03:08:01 +0000 (13:38 +1030)]
weak alias test
This adds the test I was using when testing 60d67dc8.
* testsuite/ld-elfweak/alias.c,
* testsuite/ld-elfweak/alias.dat,
* testsuite/ld-elfweak/aliasmain.c,
* testsuite/ld-elfweak/weakref1.c,
* testsuite/ld-elfweak/weakref2.c: New test.
* testsuite/ld-elfweak/elfweak.exp: Run it. Don't return on fails,
attempt other tests.
Alan Modra [Thu, 30 Nov 2017 09:16:17 +0000 (19:46 +1030)]
readonly_dynrelocs
In early October, HJ Lu added support for a number of targets to "Dump
dynamic relocation in read-only section with minfo". This extends
that support to more targets, displays the symbol involved, and splits
the existing function that sets TEXTREL into a "readonly_dynrelocs"
and "maybe_set_textrel" function. I'll need "readonly_dynrelocs" if I
ever get around to fixing "pr22374 function pointer initialization"
fails.
am33_2.0, arc, bfin, hppa64, mn10300, and nios2 fail to mark a binary
needing text relocations with DT_TEXTREL. That's not good. xtensa also
fails to do so but complains about "dangerous relocation: dynamic
relocation in read-only section" so I reckon that is fine and have
marked the test as an xfail. The other targets need maintainer
attention.
Curiously, the map file dump wasn't added for x86, so the map test
currently fail on x86. It also fails on alpha, am33_2.0, arc, bfin,
hppa64, ia64, m68k, mips, mn10300, nios2, score and vax. cris
complains with "tmpdir/textrel.o, section .rodata: relocation
R_CRIS_32 should not be used in a shared object; recompile with -fPIC"
so I've marked it as an xfail.
bfd/
* elf32-hppa.c (maybe_set_textrel): Print symbol for map file output.
* elf32-ppc.c (maybe_set_textrel): Likewise.
* elf64-ppc.c (maybe_set_textrel): Likewise.
* elf32-arm.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing..
(elf32_arm_readonly_dynrelocs): ..this.
* elf32-lm32.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf32-m32r.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf32-metag.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf32-nds32.c: Delete unnecessary forward declarations.
(readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf32-or1k.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf32-s390.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf32-sh.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf32-tic6x.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing..
(elf32_tic6x_readonly_dynrelocs): ..this.
* elf32-tilepro.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elf64-s390.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elfnn-aarch64.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing..
(aarch64_readonly_readonly_dynrelocs): ..this.
* elfnn-riscv.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elfxx-sparc.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
* elfxx-tilegx.c (readonly_dynrelocs): New function.
(maybe_set_textrel): New function, replacing old version of..
(readonly_dynrelocs): ..this.
ld/
* testsuite/ld-elf/shared.exp: Run new textrel tests.
* testsuite/ld-elf/textrel.map: New file.
* testsuite/ld-elf/textrel.rd: New file.
* testsuite/ld-elf/textrel.s: New file.
* testsuite/ld-elf/textrel.warn: New file.
Joel Brobecker [Thu, 30 Nov 2017 23:41:55 +0000 (18:41 -0500)]
New gdb.ada/repeat_dyn testcase.
This patch introduces a testcase that exercises a scenario
which used to trigger an internal-error, but no longer does:
Consider the following array:
type Small is new Integer range Ident (1) .. Ident (10);
type Table is array (1 .. 3) of Small;
A1 : Table := (3, 5, 8);
The particularity of this array is that the type of each element
is a range type whose bounds are dynamic, since they depend on
the value returned by Ident (1) and Ident (10). Trying to apply
the repeat operator ('@') on one of its elements used to yield
an internal error:
Cary Coutant [Thu, 30 Nov 2017 21:44:28 +0000 (13:44 -0800)]
Fix problem where erratum stubs are not always applied.
I checked over the results of applying --fix-cortex-a53-843419 to
a very large program (gitit) with two stub tables and thousands
of erratum fixes. I noticed that all the erratum_stubs were being
created but about 1/3 of them were being skipped over by
fix_errata_and_relocate_erratum_stubs(). By skipped over I mean
no branch relocation or adrp -> adr transformation was applied to
the erratum address, leaving the erratum_stub unreachable, and
with a branch with a 0 immediate.
The root cause of the skipped over erratum_stubs is
Erratum_stub::invalidate_erratum_stub() that is used to set
relobj_ to NULL when an erratum_stub has been processed.
Unfortunately relobj_ is used in operator<() so altering relobj
makes the results from erratum_stubs_.lower_bound() as used in
find_erratum_stubs_for_input_section() unreliable.
2017-11-30 Peter Smith <peter.smith@linaro.org>
Cary Coutant <ccoutant@gmail.com>
gold/
PR gold/20765
* aarch64.cc (Erratum_stub::invalidate_erratum_stub): Use erratum_insn_
instead of relobj_ to invalidate the stub.
(Erratum_stub::is_invalidated_erratum_stub): Likewise.
Peter Smith [Thu, 30 Nov 2017 21:30:58 +0000 (13:30 -0800)]
Fix segfault in relocate_erratum_stub on aarch64.
The fix for PR21868 (an internal error when --fix-cortex-a53-843419
is applied) has a small mistake in it. When the stub_owner section
needs an erratum fix an incorrect address for the stubs for the section
is given to relocate_erratum_stub. If we are lucky we will get a segfault;
if we aren't, an incorrect patch or data corruption is possible.
The error is visible in PR21868, but the side-effects aren't fatal.
Alan Modra [Thu, 30 Nov 2017 21:23:16 +0000 (07:53 +1030)]
PR22533, dynamic relocs generated for weak aliases
This cleans up yet more craziness with non_got_ref.
PR 22533
* elf32-hppa.c (elf32_hppa_copy_indirect_symbol): Don't do anything
special with non_got_ref for weak aliases.
(elf32_hppa_check_relocs): Tweak setting of non_got_ref.
(elf32_hppa_adjust_dynamic_symbol): When initialising weak aliases,
don't uselessly copy non_got_ref. Clear dyn_relocs instead if
strong symbol is allocated in dynbss. Tidy comments.
(elf32_hppa_relocate_section): Comment fix.
* elf32-ppc.c (ppc_elf_copy_indirect_symbol): Don't do anything
special with non_got_ref for weak aliases.
(ppc_elf_adjust_dynamic_symbol): When initialising weak aliases,
don't uselessly copy non_got_ref. Clear dyn_relocs instead if
strong symbol is allocated in dynbss. Tidy comments.
* elf64-ppc.c (ppc64_elf_copy_indirect_symbol): Don't do anything
special with non_got_ref for weak aliases.
(ppc64_elf_adjust_dynamic_symbol): When initialising weak aliases,
don't uselessly copy non_got_ref. Clear dyn_relocs instead if
strong symbol is allocated in dynbss. Tidy comments.
Ulrich Weigand [Thu, 30 Nov 2017 17:35:54 +0000 (18:35 +0100)]
[spu] Some additional test fixes
Now that the ppc64 breakpoint regression is fixed, running the
gdb.cell test suite showed a few more test case problems, caused
by tests that haven't been updated to adapt to GDB changes.
Simon Marchi [Thu, 30 Nov 2017 16:39:31 +0000 (11:39 -0500)]
Use boards/local-board.exp more
local-board.exp was introduced recently, containing the code required to
force the gdbserver boards to be non-remote (from the DejaGNU point of
view). Other board files use the same trick of forcing isremote to 0.
Instead of doing it by hand in each file, include local-board.exp.
gdb/testsuite/ChangeLog:
* boards/cc-with-tweaks.exp: Include local-board.exp instead of
setting isremote by hand.
* boards/dwarf4-gdb-index.exp: Likewise.
* boards/fission.exp: Likewise.
* boards/stabs.exp: Likewise.
Rainer Orth [Thu, 30 Nov 2017 15:05:30 +0000 (16:05 +0100)]
Remove ioctl-based procfs support on Solaris
This is the previously mentioned patch to get rid of
unstructured/ioctl-based procfs support in procfs.c. Given that support
for structured procfs was introduced in Solaris 2.6 back in 1997 and
we're just removing support for Solaris < 10, there's no point in
carrying that baggage (and tons of support for IRIX and OSF/1 as well)
around any longer.
Most of the patch should be straightforward (removing support for
!NEW_PROC_API, non-Solaris OSes and pre-Solaris 10 quirks).
Only a few points need explanations:
* <sys/syscall.h> was already included unconditionally in most places,
so there's no need to have guards in a few remaining ones.
* configure.host already obsoletes i?86-*-sysv4.2, i?86-*-sysv5, so
NEW_PROC_API detection for those in configure.ac can go.
* I'm still including <sys/procfs.h> with #define _STRUCTURED_PROC 1.
Theoretically, it would be better to include <procfs.h> on Solaris
(which includes that define), but that breaks the build over
<procfs.h> vs. gdb's "procfs.h", and doesn't exist on Linux.
* I've regenerated syscall_table[] in proc-events.c with a small script
from Solaris 10, 11.3, 11.4 <sys/syscall.h>, so there should be no
traces of older Solaris versions and other OSes left.
* prsysent_t and DYNAMIC_SYSCALLS was only used for AIX 5, but AIX
doesn't use procfs.c any longer, so all related code can go.
The patch was generated with diff -w so one can easier see changes
without being distracted by simple reindentations.
So far, it has only been compiled and smoke-tested on
amd64-pc-solaris2.1[01], sparcv9-sun-solaris2.1[01], and
x86_64-pc-linux-gnu. Certainly needs more testing (Solaris 11.3
vs. 11.4, 32-bit gdb, testsuite once I've figured out what's wrong on
Solaris 10 etc.), but it's enough to get a first impression how much
cleanup is possible here.
* proc-api.c: Remove !NEW_PROC_API support.
Remove HAVE_SYS_PROC_H and HAVE_SYS_USER_H tests.
Remove tests for macros always defined on Solaris.
* proc-events.c: Remove !NEW_PROC_API support.
Remove Remove HAVE_SYS_SYSCALL_H, HAVE_SYS_PROC_H and
HAVE_SYS_USER_H tests.
(init_syscall_table): Remove non-Solaris syscalls.
Remove tests for syscalls present on all Solaris versions.
Add missing Solaris 10+ syscalls.
(signal_table): Remove non-Solaris signals.
Remove tests for signals present on all Solaris versions.
(fault_table): Remove non-Solaris faults.
Remove tests for faults present on all Solaris versions.
* proc-flags.c: Remove !NEW_PROC_API support.
(pr_flag_table): Remove non-Solaris and pre-Solaris 7 comments.
Remove non-Solaris flags.
* proc-why.c: Remove !NEW_PROC_API support.
(pr_why_table): Remove meaningless comments.
Remove tests for reasons present on all Solaris versions.
Remove OSF/1 cases.
(proc_prettyfprint_why): Likewise.
Jan Beulich [Thu, 30 Nov 2017 10:48:13 +0000 (11:48 +0100)]
x86: derive DispN from BaseIndex
BaseIndex implies - with the exception of string instructions the
optional presence of a displacement. This is almost completely uniform
for all instructions (the sole exception being MPX ones, which don't
allow 16-bit addressing and hence Disp16), so there's no point in
explicitly stating this in the main opcode table. Drop those explict
specifications in favor of adding logic to i386-gen, shrinking the
table size quite a bit and hence making it more readable.
The opcodes/i386-tbl.h changes are due to a few cases where pointless
Disp* still hadn't been removed from their insns.
Jan Beulich [Thu, 30 Nov 2017 10:46:26 +0000 (11:46 +0100)]
x86/Intel: issue diagnostics for redundant segment override prefixes
While we shouldn't outright reject such (as was wrongly done by commit 4d36230d59 ("x86: Update segment register check in Intel syntax"), as
MASM accepts them even silently, issue (by default) a warning for such
questionable constructs.
Jan Beulich [Thu, 30 Nov 2017 10:44:27 +0000 (11:44 +0100)]
Revert "x86: Update segment register check in Intel syntax"
This reverts commit 4d36230d59903b92fbe2b53b31ed64a884860f0e.
I was committed without maintainer ack and regresses intended
functionality. A replacement will be committed shortly.
Rainer Orth [Thu, 30 Nov 2017 09:57:04 +0000 (10:57 +0100)]
Fix sol-thread.c compilation on Solaris
Building current gdb mainline with gcc 7.1 on Solaris 11.4 fails:
/vol/src/gnu/gdb/gdb/dist/gdb/sol-thread.c: In function `void _initialize_sol_thread()':
/vol/src/gnu/gdb/gdb/dist/gdb/sol-thread.c:1229:66: error: invalid conversion from `void (*)(char*, int)' to `void (*)(const char*, int)' [-fpermissive]
_("Show info on Solaris user threads."), &maintenanceinfolist);
^
In file included from /vol/src/gnu/gdb/gdb/dist/gdb/completer.h:21:0,
from /vol/src/gnu/gdb/gdb/dist/gdb/symtab.h:31,
from /vol/src/gnu/gdb/gdb/dist/gdb/language.h:26,
from /vol/src/gnu/gdb/gdb/dist/gdb/frame.h:72,
from /vol/src/gnu/gdb/gdb/dist/gdb/gdbarch.h:39,
from /vol/src/gnu/gdb/gdb/dist/gdb/defs.h:557,
from /vol/src/gnu/gdb/gdb/dist/gdb/sol-thread.c:51:
/vol/src/gnu/gdb/gdb/dist/gdb/command.h:140:33: note: initializing argument 3 of `cmd_list_element* add_cmd(const char*, command_class, void (*)(const char*, int), const char*, cmd_list_element**)'
extern struct cmd_list_element *add_cmd (const char *, enum command_class,
^~~~~~~
The following patch allows compilation to succeed on i386-pc-solaris2.11
and sparc-sun-solaris2.11.
* sol-thread.c (info_solthreads): Constify args.
Cast args to void *.
Alan Modra [Thu, 30 Nov 2017 02:40:35 +0000 (13:10 +1030)]
Fix testcase for PR ld/22471
ASFLAGS should be restored after any modification. Fixes
metag-linux +FAIL: Long branch stub
metag-linux +FAIL: Long branch stub (PIC, app)
mn10300-elf +FAIL: difference of two same-section symbols where the difference is held in another section
mn10300-elf +FAIL: relaxation and alignment directives
PR 22471
* testsuite/ld-elf/shared.exp: Save and restore ASFLAGS.
John Baldwin [Thu, 30 Nov 2017 00:28:01 +0000 (16:28 -0800)]
Define MPFR_USE_INTMAX_T so that mpfr.h assumes intmax_t is available.
mpfr.h uses a non-portable test to guess if intmax_t is available and
if API functions using intmax_t should be exposed. Define
MPFR_USE_INTMAX_T to override the non-portable test and always expose
these functions. This fixes the build on platforms where the test
guesses incorrectly.
Joel Brobecker [Wed, 29 Nov 2017 23:31:18 +0000 (18:31 -0500)]
preserve type length in ada-lang.c::to_fixed_range_type
This patch fixes a potential issue which was noticed by code inspection:
ada-lang.c::to_fixed_range_type uses gdbtypes.c::create_static_range_type
to create most of the range type, which relies on create_range_type to
do most of the work. The latter has the following piece of code which
sets the length of the range type to match the length of the index_type:
With a type like this one, create_range_type returns a type whose
size is 4 bytes, instead of 2, which is not what we we would normally
expect.
Currently, this function is only used to handle array index types,
so the length of the type actually does not matter and there should
not be any user-visible consequences of the current behavior. But
it seems best to plug this latent bug now, rather than wait for it
to surface....
gdb/ChangeLog:
* ada-lang.c (to_fixed_range_type): Make sure that the size
of the range type being returned is the same as the size
of the range type being fixed.
Pedro Alves [Wed, 29 Nov 2017 19:33:24 +0000 (19:33 +0000)]
Breakpoints in symbols with ABI tags (PR c++/19436)
Trying to set a breakpoint in a function with an ABI tag does not work
currently. E.g., debugging gdb itself, we see this with the
"string_printf" function:
(top-gdb) b string_print [TAB]
(top-gdb) b string_printf[abi:cxx11](char const*, ...) [RET]
No source file named string_printf[abi.
Make breakpoint pending on future shared library load? (y or [n])
Quoting doesn't help:
(top-gdb) b 'string_printf[abi:cxx11]'(char const*, ...)
malformed linespec error: unexpected string, "(char const*, ...)"
(top-gdb) b 'string_printf[abi:cxx11](char const*, ...)'
No source file named string_printf[abi.
Make breakpoint pending on future shared library load? (y or [n]) n
This patch fixes this, and takes it a bit further.
The actual symbol name as demangled by libiberty's demangler is really
string_printf[abi:cxx11](char const*, ...)
however, this patch makes it possible to set the breakpoint with
string_printf(char const*, ...)
too. I.e., ignoring the ABI tag.
And to match, it teaches the completer to complete the symbol name
without the ABI tag, i.e.,
(gdb) b B::function<tab>
(anonymous namespace)::B::function() B::function() const Bn::B::function()
(anonymous namespace)::B::function() const B::function_const() const
B::function() Bn::(anonymous namespace)::B::function()
(gdb) b B::function
Breakpoint 1 at 0x40072c: B::function. (6 locations)
To get back the original behavior of interpreting the function name as
a fully-qualified name, you can use the new "-qualified" (or "-q")
option/flag (added by this commit). For example:
(gdb) b -qualified B::function
B::function() B::function() const B::function_const() const
I've chosen "-qualified" / "-q" because "-f" (for "full" or
"fully-qualified") is already taken for "-function".
Note: the "-qualified" option works with both linespecs and explicit
locations. I.e., these are equivalent:
(gdb) b -q func
(gdb) b -q -f func
and so are these:
(gdb) b -q filename.cc:func
(gdb) b -q -s filename.cc -f func
(gdb) b -s filename.cc -q -f func
(gdb) b -s filename.cc -f func -q
To better understand why I consider wild matching the better default,
consider what happens when we get to the point when _all_ of GDB is
wrapped under "namespace gdb {}". I have a patch series that does
that, and when I started debugging that GDB, I immediately became
frustrated. You'd have to write "b gdb::internal_error", "b
gdb::foo", "b gdb::bar", etc. etc., which gets annoying pretty
quickly. OTOH, consider how this makes it very easy to set
breakpoints in classes wrapped in anonymous namespaces. You just
don't think of them, GDB finds the symbols for you automatically.
(At the Cauldron a couple months ago, several people told me that they
run into a similar issue when debugging other C++ projects. One
example was when debugging LLVM, which puts all its code under the
"llvm" namespace.)
Implementation-wise, what the patch does is:
- makes C++ symbol name hashing only consider the last component of
a symbol name. (so that we can look up symbol names by
last-component name only).
- adds a C++ symbol name matcher for symbol_name_match_type::WILD,
which ignores missing leading specifiers / components.
- adjusts a few preexisting testsuite tests to use "-qualified" when
they mean it.
- adds new testsuite tests.
- adds unit tests.
Grows the gdb.linespec/ tests like this:
-# of expected passes 7823
+# of expected passes 8977
gdb/ChangeLog:
2017-11-29 Pedro Alves <palves@redhat.com>
* NEWS: Mention that breakpoints on C++ functions are now set on
on all namespaces/classes by default, and mention "break
-qualified".
* ax-gdb.c (agent_command_1): Adjust to pass a
symbol_name_match_type to new_linespec_location.
* breakpoint.c (parse_breakpoint_sals): Adjust to
get_linespec_location's return type change.
(strace_marker_create_sals_from_location): Adjust to pass a
symbol_name_match_type to new_linespec_location.
(strace_marker_decode_location): Adjust to get_linespec_location's
return type change.
(strace_command): Adjust to pass a symbol_name_match_type to
new_linespec_location.
(LOCATION_HELP_STRING): Add paragraph about wildmatching, and
mention "-qualified".
* c-lang.c (cplus_language_defn): Install cp_search_name_hash.
* completer.c (explicit_location_match_type::MATCH_QUALIFIED): New
enumerator.
(complete_address_and_linespec_locations): New parameter
'match_type'. Pass it down.
(explicit_options): Add "-qualified".
(collect_explicit_location_matches): Pass the requested match type
to the linespec completers. Handle MATCH_QUALIFIED.
(location_completer): Handle "-qualified" combined with linespecs.
* cp-support.c (cp_search_name_hash): New.
(cp_symbol_name_matches_1): Implement wild matching for C++.
(cp_fq_symbol_name_matches): Reimplement.
(cp_get_symbol_name_matcher): Return different matchers depending
on the lookup name's match type.
(selftests::test_cp_symbol_name_matches): Add wild matching tests.
* cp-support.h (cp_search_name_hash): New declaration.
* dwarf2read.c
(selftests::dw2_expand_symtabs_matching::test_symbols): Add
symbols.
(test_dw2_expand_symtabs_matching_symbol): Add wild matching
tests.
* guile/scm-breakpoint.c (gdbscm_register_breakpoint_x): Adjust to
pass a symbol_name_match_type to new_linespec_location.
* linespec.c (linespec_parse_basic): Lookup function symbols using
the parser's symbol name match type.
(convert_explicit_location_to_linespec): New
symbol_name_match_type parameter. Pass it down to
find_linespec_symbols.
(convert_explicit_location_to_sals): Pass the location's name
match type to convert_explicit_location_to_linespec.
(parse_linespec): New match_type parameter. Save it in the
parser.
(linespec_parser_new): Default to symbol_name_match_type::WILD.
(linespec_complete_function): New symbol_name_match_type
parameter. Use it.
(complete_linespec_component): Pass down the parser's recorded
name match type.
(linespec_complete_label): New symbol_name_match_type parameter.
Use it.
(linespec_complete): New symbol_name_match_type parameter. Save
it in the parser and pass it down. Adjust to
get_linespec_location's prototype change.
(find_function_symbols, find_linespec_symbols): New
symbol_name_match_type parameter. Pass it down instead of
assuming symbol_name_match_type::WILD.
* linespec.h (linespec_complete, linespec_complete_function)
(linespec_complete_label): New symbol_name_match_type parameter.
* location.c (event_location::linespec_location): Now a struct
linespec_location.
(EL_LINESPEC): Adjust.
(initialize_explicit_location): Default to
symbol_name_match_type::WILD.
(new_linespec_location): New symbol_name_match_type parameter.
Record it in the location.
(get_linespec_location): Now returns a struct linespec_location.
(new_explicit_location): Also copy func_name_match_type.
(explicit_to_string_internal)
(string_to_explicit_location): Handle "-qualified".
(copy_event_location): Adjust to LINESPEC_LOCATION type change.
Copy symbol_name_match_type fields.
(event_location_deleter::operator()): Adjust to LINESPEC_LOCATION
type change.
(event_location_to_string): Adjust to LINESPEC_LOCATION type
change. Handle "-qualfied".
(string_to_explicit_location): Handle "-qualified".
(string_to_event_location_basic): New symbol_name_match_type
parameter. Pass it down.
(string_to_event_location): Handle "-qualified".
* location.h (struct linespec_location): New.
(explicit_location::func_name_match_type): New field.
(new_linespec_location): Now returns a const linespec_location *.
(string_to_event_location_basic): New symbol_name_match_type
parameter.
(explicit_completion_info::saw_explicit_location_option): New
field.
* mi/mi-cmd-break.c (mi_cmd_break_insert_1): Adjust to pass a
symbol_name_match_type to new_linespec_location.
* python/py-breakpoint.c (bppy_init): Likewise.
* python/python.c (gdbpy_decode_line): Likewise.
gdb/testsuite/ChangeLog:
2017-11-29 Pedro Alves <palves@redhat.com>
* gdb.base/langs.exp: Use -qualified.
* gdb.cp/meth-typedefs.exp: Use -qualified, and add tests without
it.
* gdb.cp/namespace.exp: Use -qualified.
* gdb.linespec/cpcompletion.exp (overload-2, fqn, fqn-2)
(overload-3, template-overload, template-ret-type, const-overload)
(const-overload-quoted, anon-ns, ambiguous-prefix): New
procedures.
(test_driver): Call them.
* gdb.cp/save-bp-qualified.cc: New.
* gdb.cp/save-bp-qualified.exp: New.
* gdb.linespec/explicit.exp: Test -qualified.
* lib/completion-support.exp (completion::explicit_opts_list): Add
"-qualified".
* lib/gdb.exp (gdb_breakpoint): Handle "qualified".
gdb/doc/ChangeLog:
2017-11-29 Pedro Alves <palves@redhat.com>
* gdb.texinfo (Linespec Locations): Document how "function" is
interpreted in C++ and Ada. Document "-qualified".
(Explicit Locations): Document how "-function" is interpreted in
C++ and Ada. Document "-qualified".
Pedro Alves [Wed, 29 Nov 2017 19:33:23 +0000 (19:33 +0000)]
Handle custom completion match prefix / LCD
A following patch will add support for wild matching for C++ symbols,
making completing on "b push_ba" on a C++ program complete to
std::vector<...>::push_back, std::string::push_back etc., like:
(gdb) b push_ba[TAB]
std::vector<...>::push_back(....)
std::string<...>::push_back(....)
Currently, we compute the "lowest common denominator" between all
completion candidates (what the input line is adjusted to) as the
common prefix of all matches. That's problematic with wild matching
as above, as then we'd end up with TAB changing the input line to
"b std::", losing the original input, like:
(gdb) b push_ba[TAB]
std::vector<...>::push_back(....)
std::string<...>::push_back(....)
(gdb) b std::
while obviously we'd want it to adjust itself to "b push_back(" instead:
(gdb) b push_ba[TAB]
std::vector<...>::push_back(....)
std::string<...>::push_back(....)
(gdb) b push_back(
This patch adds the core code necessary to support this, though
nothing really makes use of it yet in this patch.
gdb/ChangeLog:
2017-11-29 Pedro Alves <palves@redhat.com>
* ada-lang.c (ada_lookup_name_info::matches): Change type of
parameter from completion_match to completion_match_result.
Adjust.
(do_wild_match, do_full_match, ada_symbol_name_matches): Likewise.
* completer.c (completion_tracker::maybe_add_completion): Add
match_for_lcd parameter and use it.
(completion_tracker::add_completion): Likewise.
* completer.h (class completion_match_for_lcd): New class.
(completion_match_result::match_for_lcd): New field.
(completion_match_result::set_match): New method.
(completion_tracker): Add comments.
(completion_tracker::add_completion): Add match_for_lcd parameter.
(completion_tracker::reset_completion_match_result): Reset
match_for_lcd too.
(completion_tracker::maybe_add_completion): Add match_for_lcd
parameter.
(completion_tracker::m_lowest_common_denominator_unique): Extend
comments.
* cp-support.c (cp_symbol_name_matches_1)
(cp_fq_symbol_name_matches): Change type of parameter from
completion_match to completion_match_result. Adjust.
* language.c (default_symbol_name_matcher): Change type of
parameter from completion_match to completion_match_result.
Adjust.
* language.h (completion_match_for_lcd): Forward declare.
(default_symbol_name_matcher): Change type of parameter from
completion_match to completion_match_result.
* symtab.c (compare_symbol_name): Adjust.
(completion_list_add_name): Pass the match_for_lcd to the tracker.
* symtab.h (ada_lookup_name_info::matches): Change type of
parameter from completion_match to completion_match_result.
(symbol_name_matcher_ftype): Likewise, and update comments.
Pedro Alves [Wed, 29 Nov 2017 19:25:58 +0000 (19:25 +0000)]
Fix setting-breakpoints regression on PPC64 (function descriptors)
The recent-ish commit e5f25bc5d6db ('Fix "list ambiguous_variable"')
caused a serious regression on PPC64. See
<https://sourceware.org/ml/gdb-patches/2017-11/msg00666.html>.
Basically, after that patch, GDB sets breakpoints in function
descriptors instead of where the descriptors point to, which is
incorrect.
The problem is that GDB now only runs a minsym's address through
gdbarch_convert_from_func_ptr_addr if msymbol_is_text returns true.
However, if the symbol points to a function descriptor,
msymbol_is_text is false since function descriptors are in fact
outside the text section.
The fix is to also run a non-text address through
gdbarch_convert_from_func_ptr_addr, and if that detects that it was
indeed a function descriptor, treat the resulting address as a
function.
While implementing that directly in linespec.c:minsym_found (where the
bad msymbol_is_text check is) fixes the issue, I noticed that
linespec.c:add_minsym has some code that also basically needs to do
the same checks, however it's implemented differently. Also,
add_minsym is calling find_pc_sect_line on non-function symbols, which
also doesn't look right.
So I introduced msymbol_is_function, so that we have a simple place to
consider minsyms and function descriptors.
And then, the only other use of msymbol_is_text is in
find_function_alias_target, which turns out to also be incorrect.
Changing that one to use msymbol_is_function, i.e., to consider
function descriptors too fixes (on PPC64):
-FAIL: gdb.base/symbol-alias.exp: p func_alias
-FAIL: gdb.base/symbol-alias.exp: p *func_alias()
+PASS: gdb.base/symbol-alias.exp: p func_alias
+PASS: gdb.base/symbol-alias.exp: p *func_alias()
And then after that, msymbol_is_text is no longer used anywhere, so it
can be removed.
Tested on x86_64 GNU/Linux, no regressions. Tested on PPC64 GNU/Linux
and results compared to a testrun of e5f25bc5d6db^ (before the
offending commit), also no regressions. (there's a couple new FAILs
and some new symbol name matching unit tests are crashing, but that
looks unrelated).
gdb/ChangeLog:
2017-11-29 Pedro Alves <palves@redhat.com>
* linespec.c (minsym_found, add_minsym): Use msymbol_is_function.
* minsyms.c (msymbol_is_text): Delete.
(msymbol_is_function): New function.
* minsyms.h (msymbol_is_text): Delete.
(msymbol_is_function): New declaration.
* symtab.c (find_function_alias_target): Use msymbol_is_function.
projects / thirdparty / binutils-gdb.git / log
