When building with clang 15, I got this error:
CXX server.o
server.cc:2985:10: error: variable 'new_argc' set but not used [-Werror,-Wunused-but-set-variable]
int i, new_argc;
^
Remove the unused variable to eliminate the error.
Tested by rebuilding on x86_64-linux with clang 15.
We have in per_cu->set_lang this comment:
...
void set_lang (enum language lang)
{
/* We'd like to be more strict here, similar to what is done in
set_unit_type, but currently a partial unit can go from unknown to
minimal to ada to c. */
...
Fix this by not setting m_lang for partial units.
This requires us to move the m_unit_type initialization to ensure that
m_unit_type is initialized before per_cu->m_lang.
Tested on x86_64-linux, with native and target board cc-with-dwz-m.
This improves the "set scheduler-locking" documentation in the GDB
manual:
- Use a table to describe the four available modes.
- Describe "step" in terms of "on" and "off".
- Tweak the "replay" mode's description to describe replay first
instead of recording, and also mention how the mode behaves during
normal execution.
- Say what is the default mode.
Change-Id: Ie12140138b37534b7fc1d904da34f0f174aa11ce
The cu->per_cu->lang field was added to carry information from the initial
partial symtabs phase to the symtab expansion phase, for the benefit of a
particular optimization in process_imported_unit_die.
Other uses have been added, but since the first phase now has been
parallelized, those have become problematic and sources of race conditions.
Fix this by adding dwarf2_cu::lang () and using it where we can to replace
cu->per_cu->lang () with cu->lang ().
Also assert in dwarf2_cu::lang () that we're not returning language_unknown.
Tested on x86_64-linux.
When building gdb with -fsanitize=address we run into:
...
builtin_spawn gdb -nw -nx -iex set height 0 -iex set width 0 -data-directory \
build/gdb/data-directory^M
==10637==ASan runtime does not come first in initial library list; you \
should either link runtime to your application or manually preload it with \
LD_PRELOAD.^M
ERROR: GDB process no longer exists
...
Prevent the ASan runtime error by using
ASAN_OPTIONS=verify_asan_link_order=0. This makes both test-cases pass.
Tested on x86_64-linux.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29358
Add a new file tsan-suppressions.txt, to suppress the "unlock unlocked mutex"
problem in ncurses, filed in PR29328.
The file is added to the TSAN_OPTIONS in lib/gdb.exp.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29328
When building gdb with gcc 4.8.5, we run into problems due to unconditionally
using:
...
gdb_static_assert (std::is_trivially_copyable<packed>::value);
...
in gdbsupport/packed.h.
Fix this by guarding the usage with HAVE_IS_TRIVIALLY_COPYABLE.
Tested by doing a full gdb build with gcc 4.8.5.
When building gdb with -fsanitize=thread and gcc 12, and running test-case
gdb.dwarf2/dwz.exp, we run into a data race between:
...
Read of size 1 at 0x7b200000300d by thread T2:^M
#0 cutu_reader::cutu_reader(dwarf2_per_cu_data*, dwarf2_per_objfile*, \
abbrev_table*, dwarf2_cu*, bool, abbrev_cache*) gdb/dwarf2/read.c:6164 \
(gdb+0x82ec95)^M
...
and:
...
Previous write of size 1 at 0x7b200000300d by main thread:^M
#0 prepare_one_comp_unit gdb/dwarf2/read.c:23588 (gdb+0x86f973)^M
...
In other words, between:
...
if (this_cu->reading_dwo_directly)
...
and:
...
cu->per_cu->lang = pretend_language;
...
Likewise, we run into a data race between:
...
Write of size 1 at 0x7b200000300e by thread T4:
#0 process_psymtab_comp_unit gdb/dwarf2/read.c:6789 (gdb+0x830720)
...
and:
...
Previous read of size 1 at 0x7b200000300e by main thread:
#0 cutu_reader::cutu_reader(dwarf2_per_cu_data*, dwarf2_per_objfile*, \
abbrev_table*, dwarf2_cu*, bool, abbrev_cache*) gdb/dwarf2/read.c:6164 \
(gdb+0x82edab)
...
In other words, between:
...
this_cu->unit_type = DW_UT_partial;
...
and:
...
if (this_cu->reading_dwo_directly)
...
Likewise for the write to addresses_seen in cooked_indexer::check_bounds and a
read from is_dwz in dwarf2_find_containing_comp_unit for test-case
gdb.dwarf2/dw2-dir-file-name.exp and target board cc-with-dwz-m.
The problem is that the written fields are part of the same memory location as
the read fields, so executing a read and write in different threads is
undefined behavour.
Making the written fields separate memory locations, using the new
struct packed template fixes this.
The set of fields has been established experimentally to be the
minimal set to get rid of this type of -fsanitize=thread errors, but
more fields might require the same treatment.
Looking at the properties of the lang field, unlike dwarf_version it's
not available in the unit header, so it will be set the first time
during the parallel cooked index reading. The same holds for
unit_type, and likewise for addresses_seen.
dwarf2_per_cu_data::addresses_seen is moved so that the bitfields that
currently follow it can be merged in the same memory location as the
bitfields that currently precede it, for better packing.
Tested on x86_64-linux.
Co-Authored-By: Pedro Alves <pedro@palves.net>
Change-Id: Ifa94f0a2cebfae5e8f6ddc73265f05e7fd9e1532
When building gdb with -fsanitize=thread and gcc 12, and running test-case
gdb.dwarf2/dwz.exp, we run into a few data races. For example, between:
...
Write of size 1 at 0x7b200000300e by thread T4:
#0 process_psymtab_comp_unit gdb/dwarf2/read.c:6789 (gdb+0x830720)
...
and:
...
Previous read of size 1 at 0x7b200000300e by main thread:
#0 cutu_reader::cutu_reader(dwarf2_per_cu_data*, dwarf2_per_objfile*, \
abbrev_table*, dwarf2_cu*, bool, abbrev_cache*) gdb/dwarf2/read.c:6164 \
(gdb+0x82edab)
...
In other words, between:
...
this_cu->unit_type = DW_UT_partial;
...
and:
...
if (this_cu->reading_dwo_directly)
...
The problem is that the written fields are part of the same memory
location as the read fields, so executing a read and write in
different threads is undefined behavour.
Making the written fields separate memory locations, like this:
...
struct {
ENUM_BITFIELD (dwarf_unit_type) unit_type : 8;
};
...
fixes it, however that also increases the size of struct
dwarf2_per_cu_data, because it introduces padding due to alignment of
these new structs, which align on the natural alignment of the
specified type of their fields. We can fix that with
__attribute__((packed)), like so:
struct {
ENUM_BITFIELD (dwarf_unit_type) unit_type : 8 __attribute__((packed));
};
but to avoid having to write that in several places and add suitable
comments explaining how that concoction works, introduce a new struct
packed template that wraps/hides this. Instead of the above, we'll be
able to write:
packed<dwarf_unit_type, 1> unit_type;
Note that we can't change the type of dwarf_unit_type, as that is
defined in include/, and shared with other projects, some of those
written in C.
This patch just adds the struct packed type. Following patches will
make use of it. One of those patches will want to wrap a struct
packed in an std::atomic, like:
std::atomic<std::packed<language, 1>> m_lang;
so the new gdbsupport/packed.h header adds some operators to make
comparisions between that std::atomic and the type that the wrapped
struct packed wraps work, like in:
if (m_lang == language_c)
It would be possible to implement struct packed without using
__attribute__((packed)), by having it store an array of bytes of the
appropriate size instead, however that would make it less convenient
to debug GDB. The way it's implemented, printing a struct packed
variable just prints its field using its natural type, which is
particularly useful if the type is an enum. I believe that
__attribute__((packed)) is supported by all compilers that are able to
build GDB. Even a few BFD headers use on ATTRIBUTE_PACKED on external
types:
include/coff/external.h: } ATTRIBUTE_PACKED
include/coff/external.h:} ATTRIBUTE_PACKED ;
include/coff/external.h:} ATTRIBUTE_PACKED ;
include/coff/pe.h:} ATTRIBUTE_PACKED ;
include/coff/pe.h:} ATTRIBUTE_PACKED;
include/elf/external.h:} ATTRIBUTE_PACKED Elf_External_Versym;
It is not possible to build GDB with MSVC today, but if it could, that
would be one compiler that doesn't support this attribute. However,
it supports packing via pragmas, so there's a way to cross that bridge
if we ever get to it. I believe any compiler worth its salt supports
some way of packing.
In any case, the worse that happens without the attribute is that some
types become larger than ideal. Regardless, I've added a couple
static assertions to catch such compilers in action:
/* Ensure size and aligment are what we expect. */
gdb_static_assert (sizeof (packed) == Bytes);
gdb_static_assert (alignof (packed) == 1);
Change-Id: Ifa94f0a2cebfae5e8f6ddc73265f05e7fd9e1532
It might be possible to hit md_end before md_begin is called, don't
segfault if so. Also, remove a useless check.
* gas/config/tc-ppc.c (insn_calloc): Remove needless overflow
check.
(ppc_md_end): Check ppc_hash before deleting. Clear ppc_hash.
The bug testcase uses an output section named .rel or .rela which has
input .data sections mapped to it. The input .data section has
relocations. When counting output relocations SHT_REL and SHT_RELA
section reloc_count is ignored, with the justification that reloc
sections themselves can't have relocations and some backends use
reloc_count in reloc sections. However, the test wrongly used the
output section type (which normally would match input section type).
Fix that. Note that it is arguably wrong for ld to leave the output
.rel/.rela section type as SHT_REL/SHT_RELA when non-empty non-reloc
sections are written to it, but I'm not going to change that since it
might be useful to hand-craft relocs in a data section that is then
written to a SHT_REL/SHT_RELA output section.
PR 29355
* elflink.c (bfd_elf_final_link): Use input section type rather
than output section type to determine whether to exclude using
reloc_count from that section.
For csky arch, the correspondence between Dwarf registers and GDB
registers are as follows:
dwarf regnos 0~31 ==> gdb regs r0~r31
dwarf regno CSKY_HI_REGNUM(36) ==> gdb reg hi
dwarf regno CSKY_LO_REGNUM(37) ==> gdb reg hi
dwarf regno CSKY_PC_REGNUM(72) ==> gdb reg pc
dwarf regnos FV_PSEUDO_REGNO_FIRST(74)~FV_PSEUDO_REGNO_LAST(201)
==>
gdb regs s0~s127 (pseudo regs for float and vector regs)
other dwarf regnos have no corresponding gdb regs to them.
[Note: the testcased added by this commit depends on
https://sourceware.org/pipermail/gdb-patches/2022-June/190259.html,
otherwise GDB just crashes when detaching the core]
Currently, in MI, =thread-created are always emitted, like:
=thread-group-started,id="i1",pid="195680"
=thread-created,id="1",group-id="i1"
...
but on teardown, if the target uses exit_inferior_silent, then you'll
only see the inferior exit notification (thread-group-exited), no
notification for threads.
The core target is one of the few targets that use
exit_inferior_silent. Here's an example session:
-target-select core $corefile
=thread-group-started,id="i1",pid="195680"
=thread-created,id="1",group-id="i1"
...
^connected,frame=....
(gdb)
-target-detach
=thread-group-exited,id="i1"
^done
(gdb)
This imbalance of emitting =thread-created but then not =thread-exited
seems off to me. (And, it complicates changes I want to do to
centralize emitting thread exit notifications for the CLI, which is
why I'm looking at this.)
And then, since most other targets use exit_inferior instead of
exit_inferior_silent, MI is already emitting =thread-exited
notifications when tearing down an inferior, for most targets.
This commit makes MI always emit the =thread-exited notifications,
even for exit_inferior_silent.
Afterwards, when debugging a core, MI outputs:
(gdb)
-target-detach
=thread-exited,id="1",group-id="i1" << new line
=thread-group-exited,id="i1"
^done
(gdb)
Surprisingly, there's no MI testcase debugging a core file. This
commit adds the first.
Change-Id: I5100501a46f07b6bbad3e04d120c2562a51c93a4
After loading a core file, you're supposed to be able to use "detach"
to unload the core file. That unfortunately regressed starting with
GDB 11, with these commits:
1192f124a3 - gdb: generalize commit_resume, avoid commit-resuming when threads have pending statuses
408f66864a - detach in all-stop with threads running
resulting in a GDB crash:
...
Thread 1 "gdb" received signal SIGSEGV, Segmentation fault.
0x0000555555e842bf in maybe_set_commit_resumed_all_targets () at ../../src/gdb/infrun.c:2899
2899 if (proc_target->commit_resumed_state)
(top-gdb) bt
#0 0x0000555555e842bf in maybe_set_commit_resumed_all_targets () at ../../src/gdb/infrun.c:2899
#1 0x0000555555e848bf in scoped_disable_commit_resumed::reset (this=0x7fffffffd440) at ../../src/gdb/infrun.c:3023
#2 0x0000555555e84a0c in scoped_disable_commit_resumed::reset_and_commit (this=0x7fffffffd440) at ../../src/gdb/infrun.c:3049
#3 0x0000555555e739cd in detach_command (args=0x0, from_tty=1) at ../../src/gdb/infcmd.c:2791
#4 0x0000555555c0ba46 in do_simple_func (args=0x0, from_tty=1, c=0x55555662a600) at ../../src/gdb/cli/cli-decode.c:95
#5 0x0000555555c112b0 in cmd_func (cmd=0x55555662a600, args=0x0, from_tty=1) at ../../src/gdb/cli/cli-decode.c:2514
#6 0x0000555556173b1f in execute_command (p=0x5555565c5916 "", from_tty=1) at ../../src/gdb/top.c:699
The code that crashes looks like:
static void
maybe_set_commit_resumed_all_targets ()
{
scoped_restore_current_thread restore_thread;
for (inferior *inf : all_non_exited_inferiors ())
{
process_stratum_target *proc_target = inf->process_target ();
if (proc_target->commit_resumed_state)
^^^^^^^^^^^
With 'proc_target' above being null. all_non_exited_inferiors filters
out inferiors that have pid==0. We get here at the end of
detach_command, after core_target::detach has already run, at which
point the inferior _should_ have pid==0 and no process target. It is
clear it no longer has a process target, but, it still has a pid!=0
somehow.
The reason the inferior still has pid!=0, is that core_target::detach
just unpushes, and relies on core_target::close to actually do the
getting rid of the core and exiting the inferior. The problem with
that is that detach_command grabs an extra strong reference to the
process stratum target, so the unpush_target inside
core_target::detach doesn't actually result in a call to
core_target::close.
Fix this my moving the cleaning up the core inferior to a shared
routine called by both core_target::close and core_target::detach. We
still need to cleanup the inferior from within core_file::close
because there are paths to it that want to get rid of the core without
going through detach. E.g., "core-file" -> "run".
This commit includes a new test added to gdb.base/corefile.exp to
cover the "core-file core" -> "detach" scenario.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29275
Change-Id: Ic42bdd03182166b19f598428b0dbc2ce6f67c893
In the description of stop replies, where the "thread" register is
explained, and where the fork and vfork stop reasons are detailed,
there are references to "@var{thread-id}", but such variable does not
actually exist. This commit fixes it.
Change-Id: If679944aaf15f6f64aabe506339a9e7667864cab
The -msign-return-address switch has been dropped from GCC, but some
older compiler may still support it. Make sure we try both
-msign-return-address and -mbranch-protection before bailing out when
running gdb.arch/aarch64-pauth.exp.
This commit extends GDB to make use of libopcodes styling support
where available, currently this is just i386 based architectures, and
RISC-V.
For architectures that don't support styling using libopcodes GDB will
fall back to using the Python Pygments package, when the package is
available.
The new libopcodes based styling has the disassembler identify parts
of the disassembled instruction, e.g. registers, immediates,
mnemonics, etc, and can style these components differently.
Additionally, as the styling is now done in GDB we can add settings to
allow the user to configure which colours are used right from the GDB
CLI.
There's some new maintenance commands:
maintenance set libopcodes-styling enabled on|off
maintenance show libopcodes-styling
These can be used to manually disable use of libopcodes styling. This
is a maintenance command as it's not anticipated that a user should
need to do this. But, this could be useful for testing, or, in some
rare cases, a user might want to override the Python hook used for
disassembler styling, and then disable libopcode styling so that GDB
falls back to using Python. Right now I would consider this second
use case a rare situation, which is why I think a maintenance command
is appropriate.
When libopcodes is being used for styling then the user can make use
of the following new styles:
set/show style disassembler comment
set/show style disassembler immediate
set/show style disassembler mnemonic
set/show style disassembler register
The disassembler also makes use of the 'address' and 'function'
styles to style some parts of the disassembler output. I have also
added the following aliases though:
set/show style disassembler address
set/show style disassembler symbol
these are aliases for:
set/show style address
set/show style function
respectively, and exist to make it easier for users to discover
disassembler related style settings. The 'address' style is used to
style numeric addresses in the disassembler output, while the 'symbol'
or 'function' style is used to style the names of symbols in
disassembler output.
As not every architecture supports libopcodes styling, the maintenance
setting 'libopcodes-styling enabled' has an "auto-off" type behaviour.
Consider this GDB session:
(gdb) show architecture
The target architecture is set to "auto" (currently "i386:x86-64").
(gdb) maintenance show libopcodes-styling enabled
Use of libopcodes styling support is "on".
the setting defaults to "on" for architectures that support libopcodes
based styling.
(gdb) set architecture sparc
The target architecture is set to "sparc".
(gdb) maintenance show libopcodes-styling enabled
Use of libopcodes styling support is "off" (not supported on architecture "sparc")
the setting will show as "off" if the user switches to an architecture
that doesn't support libopcodes styling. The underlying setting is
still "on" at this point though, if the user switches back to
i386:x86-64 then the setting would go back to being "on".
(gdb) maintenance set libopcodes-styling enabled off
(gdb) maintenance show libopcodes-styling enabled
Use of libopcodes styling support is "off".
now the setting is "off" for everyone, even if the user switches back
to i386:x86-64 the setting will still show as "off".
(gdb) maintenance set libopcodes-styling enabled on
Use of libopcodes styling not supported on architecture "sparc".
(gdb) maintenance show libopcodes-styling enabled
Use of libopcodes styling support is "off".
attempting to switch the setting "on" for an unsupported architecture
will give an error, and the setting will remain "off".
(gdb) set architecture auto
The target architecture is set to "auto" (currently "i386:x86-64").
(gdb) maintenance show libopcodes-styling enabled
Use of libopcodes styling support is "off".
(gdb) maintenance set libopcodes-styling enabled on
(gdb) maintenance show libopcodes-styling enabled
Use of libopcodes styling support is "on".
the user will need to switch back to a supported architecture before
they can one again turn this setting "on".
The gdb_disassemble_info class is a wrapper around the libopcodes
disassemble_info struct. The 'stream' field of disassemble_info is
passed as an argument to the fprintf_func and fprintf_styled_func
callbacks when the disassembler wants to print anything.
Previously, GDB would store a pointer to a ui_file object in the
'stream' field, then, when the disassembler wanted to print anything,
the content would be written to the ui_file object. An example of an
fprintf_func callback, from gdb/disasm.c is:
int
gdb_disassembler::dis_asm_fprintf (void *stream, const char *format, ...)
{
/* Write output to STREAM here. */
}
This is fine, but has one limitation, within the print callbacks we
only have access to STREAM, we can't access any additional state
stored within the gdb_disassemble_info object.
Right now this isn't a problem, but in a future commit this will
become an issue, how we style the output being written to STREAM will
depend on the state of the gdb_disassemble_info object, and this state
might need to be updated, depending on what is being printed.
In this commit I propose changing the 'stream' field of the
disassemble_info to carry a pointer to the gdb_disassemble_info
sub-class, rather than the stream itself.
We then have the two sub-classes of gdb_disassemble_info to consider,
the gdb_non_printing_disassembler class never cared about the stream,
previously, for this class, the stream was nullptr. With the change
to make stream be a gdb_disassemble_info pointer, no further updates
are needed for gdb_non_printing_disassembler.
The other sub-class is gdb_printing_disassembler. In this case the
sub-class now carries around a pointer to the stream object. The
print callbacks are updated to cast the incoming stream object back to
a gdb_printing_disassembler, and then extract the stream.
This is purely a refactoring commit. A later commit will add
additional state to the gdb_printing_disassembler, and update the
print callbacks to access this state.
There should be no user visible changes after this commit.
With gdb build with -fsanitize=thread and test-case
gdb.dwarf2/dw4-sig-types.exp and target board cc-with-dwz-m we run into a data
race between:
...
Write of size 4 at 0x7b2800002268 by thread T4:^M
#0 cutu_reader::cutu_reader(dwarf2_per_cu_data*, dwarf2_per_objfile*, \
abbrev_table*, dwarf2_cu*, bool, abbrev_cache*) gdb/dwarf2/read.c:6236 \
(gdb+0x82f525)^M
...
and this read:
...
Previous read of size 4 at 0x7b2800002268 by thread T1:^M
#0 dwarf2_find_containing_comp_unit gdb/dwarf2/read.c:23444 \
(gdb+0x86e22e)^M
...
In other words, between this write:
...
this_cu->length = cu->header.get_length ();
...
and this read:
...
&& mid_cu->sect_off + mid_cu->length > sect_off))
...
of per_cu->length.
Fix this similar to the per_cu->dwarf_version case, by only setting it if
needed, and otherwise verifying that the same value is used. [ Note that the
same code is already present in the other cutu_reader constructor. ]
Move this logic into into a member function set_length to make sure it's used
consistenly, and make the field private in order to enforce access through the
member functions, and rename it to m_length.
This exposes (running test-case gdb.dwarf2/fission-reread.exp) that in
fill_in_sig_entry_from_dwo_entry, the m_length field is overwritten.
For now, allow for that exception.
While we're at it, make sure that the length is set before read.
Tested on x86_64-linux.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=29344
There's a spot in read_comp_units_from_section where we explictly use
initial_length_size to get the total length:
...
this_cu->length = cu_header.length + cu_header.initial_length_size;
...
Instead, just use cu_header.get_length ().
Tested on x86_64-linux.
Commit 736918239b ("gdb: LoongArch: add orig_a0 into register set")
introduced orig_a0, similar processing needs to be done in gdbserver.
At the same time, add orig_a0 related comments.
Signed-off-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Move "enum loongarch_regnum" to gdb/arch/loongarch.h so that the
macro definitions can be used in gdbserver/linux-loongarch-low.cc
to simplify the code.
Signed-off-by: Youling Tang <tangyouling@loongson.cn>
Signed-off-by: Tiezhu Yang <yangtiezhu@loongson.cn>
Cleaning up the subsym_hash memory is a real pain. Keys and values
entered into the table are quite diverse. In some cases the key is
allocated and thus needs to be freed, in others the key is a const
string. Values are similar, and in some cases not even a string.
Tidy this by inserting a new subsym_ent_t that describes key/value
type. This meant the math_hash table was no longer needed. The patch
also tidies how math functions are called, those that are supposed to
return int now no longer return their value in a float.
* config/tc-tic54x.c (math_hash): Delete.
(subsym_proc_entry): Move earlier, make proc a union, merge with..
(math_proc_entry): ..this. Delete type.
(math_procs): Merge into subsym_procs.
(subsym_ent_t): New. Use this type in subsym_hash..
(stag_add_field_symbols, tic54x_var, tic54x_macro_info): ..here..
(md_begin, subsym_create_or_replace, subsym_lookup): ..and here..
(subsym_substitute): ..and here. Adjust subsym_proc_entry
function calls. Free replacement when not returned.
(subsym_get_arg): Adjust subsym_lookup.
(free_subsym_ent, subsym_htab_create ): New functions, use when
creating subsym_hash.
(free_local_label_ent, local_label_htab_create): Similarly.
(tic54x_remove_local_label): Delete.
(tic54x_clear_local_labels): Simplify.
(tic54x_asg): Use notes obstack to dup strings.
(tic54x_eval): Likewise.
(subsym_ismember): Likewise.
(math_cvi, math_int, math_sgn): Return int.
(tic54x_macro_start): Decrement macro_level before calling as_fatal.
(tic54x_md_end): New function.
* config/tc-tic54h.c (tic54x_md_end): Declare.
(md_end): Define.
Using notes_calloc means all of the string hash table memory should
now be freed before gas exits, even though htab_delete isn't called.
This also means that the hash table free_f and del_f must be NULL,
because freeing notes obstack memory results in all more recently
allocated notes memory being freed too. So hash table resizing won't
free any memory, and will be a little faster. Also, htab_delete won't
do anything (and be quick about it).
Since htab_traverse can also resize hash tables (to make another
traversal faster if the table is largely empty), stop that happening
when only one traversal is done.
* as.h: Reorder hash.h after symbols.h for notes_calloc decl.
* hash.h (str_htab_create): Use notes_calloc. Do not free.
* symbols.c (resolve_local_symbol_values): Don't resize
during hash table traversal.
* config/obj-elf.c (elf_frob_file_after_relocs): Likewise.
* config/tc-ia64.c (ia64_adjust_symtab, ia64_frob_file): Likewise.
* config/tc-nds32.c (nds32_elf_analysis_relax_hint): Likewise.
This allocates entries added to the string hash tables on the notes
obstack, so that at least those do not leak. A followup patch will
switch over the str_hash allocation to notes_calloc, which is why I
haven't implemented deleting all the target string hash tables.
* config/obj-coff-seh.c (get_pxdata_name, alloc_pxdata_item): Use
notes obstack for string hash table entries.
* config/tc-alpha.c (get_alpha_reloc_tag, md_begin): Likewise.
* config/tc-h8300.c (md_begin): Likewise.
* config/tc-ia64.c (dot_rot, dot_pred_rel, dot_alias): Likewise.
* config/tc-nds32.c (nds32_relax_hint): Likewise.
* config/tc-riscv.c (riscv_init_csr_hash): Likewise.
* config/tc-score.c (s3_insert_reg): Likewise.
(s3_build_score_ops_hsh, s3_build_dependency_insn_hsh): Likewise.
* config/tc-score7.c (s7_build_score_ops_hsh): Likewise.
(s7_build_dependency_insn_hsh): Likewise.
* config/tc-tic4x.c (tic4x_asg): Likewise.
The arc opcode hash table has entries that have a realloc'd field.
This doesn't lend itself to obstack allocation, so freeing must be
done with a purpose built hashtab del_f.
* config/tc-arc.c (arc_opcode_free): New function.
(md_begin): Pass the above as del_f to htab_create_alloc.
(arc_md_end): New function.
* config/tc-arc.h (arc_md_end): Declare.
(md_end): Define.
This tidies memory used by the two x86 gas string hash tables before
exiting. I'm using a two-pronged approach, firstly the obvious call
to htab_delete plus telling the libiberty/hashtab.c infrastructure to
free tuples generated by str_hash_insert, and secondly putting the x86
core_optab memory on the notes obstack. It would be possible to free
core_optab memory by using a custom hash table del_f on x86, as I do
for arc, but a later patch will move all the string hash memory to the
notes obstack.
* config/tc-i386.c (md_begin): Use notes_alloc for core_optab.
(386_md_end): New function.
* config/tc-i386.h (386_md_end): Declare.
(md_end): Define.
* hash.h (str_htab_create): Pass free as del_f.
Only inline functions should be defined in hash.h, there's no benefit
in having multiple copies of hash_string_tuple and eq_string_tuple.
Also, use the table alloc_f when allocating tuples to be stored, so
that these functions are usable with different memory allocation
strategies.
* hash.h (struct string_tuple, string_tuple_t): Move earlier.
(string_tuple_alloc): Add table param, allocate using table alloc_f.
(str_hash_insert): Adjust to suit. Call table->free_f when
entry is not used.
(hash_string_tuple, eq_string_tuple): Move to..
* hash.c: ..here.
Currently md_end is typically used for some final actions rather than
freeing memory like other *_end functions. Rename it to md_finish,
and rename target implementation. The renaming of target functions
makes it possible to find them all with "grep md_finish",
eg. md_mips_end is renamed to mips_md_finish, not md_mips_finish.
This patch leaves a number of md_end functions unchanged, those that
either do nothing or deallocate memory, and calls them late.
The idea here is that target maintainers implement md_end functions to
tidy memory, if anyone cares. Freeing persistent memory in gas is
not at all important, except that it can hide more important memory
leaks, those that happen once per some frequent gas operation, amongst
these unimportant memory leaks.
* as.c (main): Rename md_end to md_finish.
* config/tc-alpha.c, * config/tc-alpha.h,
* config/tc-arc.c, * config/tc-arc.h,
* config/tc-arm.c, * config/tc-arm.h,
* config/tc-csky.c, * config/tc-csky.h,
* config/tc-ia64.c, * config/tc-ia64.h,
* config/tc-mcore.c, * config/tc-mcore.h,
* config/tc-mips.c, * config/tc-mips.h,
* config/tc-mmix.c, * config/tc-mmix.h,
* config/tc-msp430.c, * config/tc-msp430.h,
* config/tc-nds32.c, * config/tc-nds32.h,
* config/tc-ppc.c, * config/tc-ppc.h,
* config/tc-pru.c, * config/tc-pru.h,
* config/tc-riscv.c, * config/tc-riscv.h,
* config/tc-s390.c, * config/tc-s390.h,
* config/tc-sparc.c, * config/tc-sparc.h,
* config/tc-tic4x.c, * config/tc-tic4x.h,
* config/tc-tic6x.c, * config/tc-tic6x.h,
* config/tc-v850.c, * config/tc-v850.h,
* config/tc-xtensa.c, * config/tc-xtensa.h,
* config/tc-z80.c, * config/tc-z80.h: Similarly.
* output-file.c (output_file_close): Call md_end.
So that the notes obstack can be used for persistent storage in
parse_args.
* as.c (parse_args): Use notes_alloc and notes_strdup.
(free_notes): New function.
(main): Init notes obstack, and arrange to be freed on exit.
* read.c (read_begin): Don't init notes obstack.
(read_end): Free cond_obstack.
* subsegs.c (subsegs_end): Don't free cond_obstack or notes.
itbl_files seems to be debug code. Get rid of it.
* as.c (struct itbl_file_list): Delete.
(itbl_files): Delete.
(parse_args): Don't keep itbl_files list.
Use notes obstack for dwcfi_hash entries, and free table. Freeing the
table makes memory checkers complain more about "definitely lost"
memory as we've moved some from the "still reachable" category.
That will be fixed with a later patch.
* dw2gencfi.c (get_debugseg_name): Allocate on notes obstack.
(alloc_debugseg_item): Likewise.
(dwcfi_hash_find_or_make): Adjust failure path free.
(cfi_finish): Delete dwfci_hash.
Another case of duplicated hash.h code, the only minor difference
being that macro->format_hash was created with 7 entries vs. str_hash
with 16 entries.
* macro.c (macro_init, define_macro): Use str_htab_create.
(do_formals, define_macro, macro_expand_body): Use str_hash_insert
(macro_expand_body): Use str_hash_find and str_hash_delete.
(delete_macro): Likewise.
(sub_actual, macro_expand, check_macro): Use str_hash_find.
(expand_irp): Use str_htab_create and str_hash_insert.
* macro.h (struct macro_struct): Tidy.
(struct macro_hash_entry, macro_hash_entry_t, hash_macro_entry),
(eq_macro_entry, macro_entry_alloc, macro_entry_find),
(struct formal_hash_entry, formal_hash_entry_t),
(hash_formal_entry, eq_formal_entry, formal_entry_alloc),
(formal_entry_find): Delete.
* config/tc-iq2000.c (iq2000_add_macro): Use str_htab_create
and str_hash_insert.
po_hash code duplicates the str_hash code in hash.h for no good reason.
* read.c (struct po_entry, po_entry_t): Delete.
(hash_po_entry, eq_po_entry, po_entry_alloc, po_entry_find): Delete.
(pop_insert): Use str_hash_insert.
(pobegin): Use str_htab_create.
(read_a_source_file, s_macro): Use str_hash_find.
read_symbol_name mallocs the string it returns. Free it when done.
* read.c (read_symbol_name): Free name on error path.
* config/tc-ppc.c (ppc_GNU_visibility): Free name returned from
read_symbol_name.
(ppc_extern, ppc_globl, ppc_weak): Likewise.
This is mostly a tidy with the aim of being able to free
out_file_name, but it does fix a possible attempt to unlink the output
file twice (not that that matters).
* as.h (keep_it): New global.
* as.c (keep_it): Delete.
(close_output_file): Delete, merged into..
* output-file.c (output_file_close): ..here. Delete parameter.
* output-file.h (output_file_close): Update prototype.
