PR cli/17716 notes that it is difficult to C-c (or "q" at a pagination
prompt) while backtracing using a frame filter. One reason for this
is that many places in py-framefilter.c use RETURN_MASK_ALL in a
try/catch.
This patch changes these spots to use RETURN_MASK_ERROR instead. This
is safe to do because this entire file is exception safe now.
gdb/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
PR cli/17716:
* python/py-framefilter.c (py_print_type, py_print_value)
(enumerate_args, py_print_args, gdbpy_apply_frame_filter): Use
RETURN_MASK_ERROR.
This patch removes the last bit of manual resource management from
py-framefilter.c. This will be useful in the next patch.
gdb/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
* python/py-framefilter.c (enumerate_args): Use
gdb::unique_xmalloc_ptr.
While looking at the frame filter code, I noticed that
EXT_LANG_BT_COMPLETED is not really needed. Semantically there is no
difference between the "completed" and "ok" results. So, this patch
removes this constant.
gdb/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
* python/py-framefilter.c (py_print_frame): Return
EXT_LANG_BT_OK.
(gdbpy_apply_frame_filter): Update comment.
* extension.h (enum ext_lang_bt_status) <EXT_LANG_BT_COMPLETED>:
Remove.
<EXT_LANG_BT_NO_FILTERS>: Change value.
When a frame filter elides some frames, they are still printed by
"bt", indented a few spaces. PR backtrace/15582 notes that it would
be nice for users if elided frames could simply be dropped. This
patch adds this capability.
gdb/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
PR backtrace/15582:
* stack.c (backtrace_command): Parse "hide" argument.
* python/py-framefilter.c (py_print_frame): Handle PRINT_HIDE.
* extension.h (enum frame_filter_flags) <PRINT_HIDE>: New
constant.
gdb/doc/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
PR backtrace/15582:
* gdb.texinfo (Backtrace): Mention "hide" argument.
gdb/testsuite/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
PR backtrace/15582:
* gdb.python/py-framefilter.exp: Add "bt hide" test.
The next patch will add more flags to backtrace_command_1; and rather
than add another boolean argument, this patch changes it to accept a
flags value.
gdb/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
* stack.c (backtrace_command_1): Remove "show_locals" parameter,
add "flags".
(backtrace_command): Remove "fulltrace", add "flags".
The backtrace command has peculiar command-line parsing. In
particular, it splits the command line, then loops over the arguments.
If it sees a word it recognizes, like "full", it effectively drops
this word from the argument vector. Then, it pastes together the
remaining arguments, passing them on to backtrace_command_1, which in
turn passes the resulting string to parse_and_eval_long.
The documentation doesn't mention the parse_and_eval_long at all, so
it is a bit of a hidden feature that you can "bt 3*2". The strange
algorithm above also means you can "bt 3 * no-filters 2" and get 6
frames...
This patch changes backtrace's command line parsing to be a bit more
rational. Now, special words like "full" are only recognized at the
start of the command.
This also updates the documentation to describe the various bt options
individually.
gdb/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
* stack.c (backtrace_command): Rewrite command line parsing.
gdb/doc/ChangeLog
2018-03-26 Tom Tromey <tom@tromey.com>
* gdb.texinfo (Backtrace): Describe options individually.
While moving things around, I stumbled on filename_seen_cache being
re-defined, because filename-seen-cache.h doesn't have an include guard.
gdb/ChangeLog:
* filename-seen-cache.h: Add include guard.
In PR 22868, two IR files provide conflicting visibility for a symbol.
When a def with PROTECTED visibility is seen after a def with DEFAULT
visibility, gold does not override the visibility. Later, if the
replacement object define the symbol with DEFAULT visibility, the symbol
remains DEFAULT. This was caused by a recent change to allow multiply-defined
absolute symbols, combined with the fact that the plugin framework was using
SHN_ABS as the section index for placeholder symbols. The solution is to
use a real (but arbitrary) section index.
gold/
PR gold/22868
* plugin.cc (Sized_pluginobj::do_add_symbols): Use a real section
index instead of SHN_ABS for defined symbols.
* testsuite/Makefile.am (plugin_pr22868): New test case.
* testsuite/Makefile.in: Regenerate
* testsuite/plugin_pr22868.sh: New test script.
* testsuite/plugin_pr22868_a.c: New source file.
* testsuite/plugin_pr22868_b.c: New source file.
Buildbot pointed out a failiure in windows-nat.c:
../../binutils-gdb/gdb/windows-nat.c:582:10: error: using typedef-name 'section_addr_info' after 'struct'
struct section_addr_info *addrs;
^~~~~~~~~~~~~~~~~
In file included from ../../binutils-gdb/gdb/windows-nat.c:49:0:
../../binutils-gdb/gdb/symfile.h:75:37: note: 'section_addr_info' has a previous declaration here
typedef std::vector<other_sections> section_addr_info;
^~~~~~~~~~~~~~~~~
A recursive grep of the sources for "struct section_addr_info" reveals one
additional reference in a comment. In both cases, this patch simply removes
the struct keyword.
gdb/ChangeLog:
* symfile.c (place_section): Remove "struct" from section_addr_info
in comment.
* windows-nat.c (struct safe_symbol_file_add_args) <addrs>: Remove
"struct" keyword from section_addr_info.
I noticed that in evaluate_funcall, where we handle
OP_VAR_MSYM_VALUE/OP_VAR_VALUE to figure out the symbol's name gets
the minimal_symbol/symbol backwards. Happens to be harmless in
practice because the symbol name is recorded in the common initial
sequence (in the general_symbol_info field).
gdb/ChangeLog:
2018-03-25 Pedro Alves <palves@redhat.com>
* eval.c (evaluate_funcall): Swap OP_VAR_MSYM_VALUE/OP_VAR_VALUE
if then/else bodies in var_func_name extraction.
The TLSDESC entry in a lazy procedure linkage table is called indirectly
with "callq *(%rax)". This patch adds an ENDBR64 to support indirect
branch tracking in Intel CET. The TLSDESC PLT entry now looks like:
0xf3, 0x0f, 0x1e, 0xfa, /* endbr64 */
0xff, 0x35, 8, 0, 0, 0, /* pushq GOT+8(%rip) */
0xff, 0x25, 16, 0, 0, 0 /* jmpq *GOT+TDG(%rip) */
The BND prefix isn't needed since MPX isn't used for TLSDESC.
bfd/
PR ld/23000
* elf64-x86-64.c (elf_x86_64_finish_dynamic_sections): Add
ENDBR64 to the TLSDESC PLT entry.
ld/
PR ld/23000
* testsuite/ld-x86-64/tlsdesc.pd: Updated.
TLS variables can't be resolved on aarch64-linux-gnu
Running the test case with upstream gdb shows two failures:
(1) Receiving different error messages when printing TLS variable before
program runs - because the ARM compiler does not emit dwarf attribute
DW_AT_location for TLS, the result is expected and the baseline may
need to be changed for aarch64.
(2) Using "info address" command on C++ static TLS object resulted in
"symbol unresolved" error - below is a snippet from the test case:
class K {
public:
static __thread int another_thread_local;
};
__thread int K::another_thread_local;
(gdb) info address K::another_thread_local
Symbol "K::another_thread_local" is unresolved.
This patch contains fix for (2).
Function info_address_command() handles the "info address" command and
calls lookup_minimal_symbol_and_objfile() to find sym's symbol entry in
mininal symbol table if SYMBOL_COMPUTED_OPS (sym) is false. Problem is
that function lookup_minimal_symbol_and_objfile() only looked up an
objfile's minsym ordinary hash table, not its demangled hash table, which
was the reason why the C++ name was not found.
The fix is to call lookup_minimal_symbol(), which already looks up entries
in both minsym's hash tables, to find names when traversing the object file
list in lookup_minimal_symbol_and_objfile().
Tested in both aarch64-linux-gnu and amd64-linux-gnu. No regressions.
c++/22968 involves the inability of ptype to find a type definition for
a type defined inside another type. I recently added some additional
support for nested type definitions, but I apparently overlooked psymbols.
The user reports that using -readnow fixes the problem:
$ gdb 22968 -ex "ptype Outer::Inner"
There is no field named Inner
$ gdb -readnow 22968 -ex "ptype Outer::Inner"
type = struct Outer::Inner {
<no data field>
}
We clearly did not find a psymbol for Outer::Inner because it was located
in another CU. This patch addresses this problem by scanning structs
for additional psymbols. Rust is already doing this.
With this patch, the identical result to "-readnow" is given (without
using `-readnow', of course).
gdb/ChangeLog:
PR c++/22968
* dwarf2read.c (scan_partial_symbols): Scan structs/classes for
nested type definitions for C++, too.
gdb/testsuite/ChangeLog:
PR c++/22968
* gdb.cp/subtypes.exp: New file.
* gdb.cp/subtypes.h: New file.
* gdb.cp/subtypes.cc: New file.
* gdb.cp/subtypes-2.cc: New file.
With --debug=plugin, gold will log all plugin actions to a log file,
and make copies of plugin-supplied replacement files in a temporary
directory created under the current working directory.
gold/
* debug.h (DEBUG_PLUGIN): New constant.
(DEBUG_ALL): Add DEBUG_PLUGIN.
(debug_string_to_enum): Likewise.
* plugin.cc (make_sized_plugin_object): Add filename parameter.
(Plugin_recorder): New class.
(Plugin_manager::~Plugin_manager): Delete recorder_.
(Plugin_manager::load_plugins): Create and initialize recorder_.
(Plugin_manager::claim_file): Record claimed and unclaimed files.
(Plugin_manager::make_plugin_object): Use object name as name for
plugin object, if available.
(Plugin_manager::add_input_file): Record replacement files.
(Sized_pluginobj::do_add_symbols): Record plugin symbols.
(Plugin_finish::run): Call Plugin_recorder::finish().
(make_sized_plugin_object): Add filename parameter and pass to
Sized_pluginobj constructor.
* plugin.h (Plugin::filename): New method.
(Plugin::recorder): New method.
(Plugin::recorder_): New data member.
This changes machoread.c to use std::vector rather than VEC. This
allows removing some cleanups.
Regression tested by the buildbot, though I don't think anything
actually tests macho reading.
gdb/ChangeLog
2018-03-23 Tom Tromey <tom@tromey.com>
* machoread.c (struct oso_el): Add a constructor. Don't define as
a typedef.
(macho_register_oso): Remove.
(macho_symtab_read): Take a std::vector.
(oso_el_compare_name): Now a std::sort comparator.
(macho_symfile_read_all_oso): Take a std::vector.
(macho_symfile_read): Use std::vector. Remove cleanups.
* readelf.c (get_symbol_for_build_attribute): Skip ARM mapping
symbols.
(print_gnu_build_attribute_description): If no file start symbol
could be found, look for one two bytes into the file.
* testsuite/binutils-all/note-4-64.s: Set the address of the file
start symbol to two bytes into the file.
* testsuite/binutils-all/note-4-32.s: Likewise.
This test starts up and confirms that $xmm0 has the value 0, it then
modifies $xmm0 (in the inferior) and confirms that the new value can
be read (in GDB).
On some machines I was noticing that this test would occasionally
fail, and on investigation I believe that the reason for this is that
the test is linked as a dynamically linked executable and makes use of
the system libraries during startup. The reason that this causes
problems is that both the runtime linker and the startup code run
before main can, and do (on at least some platforms) make use of the
XMM registers.
In this commit I modify the test program slightly to allow it to be
linked statically, without using the startup libraries. Now by the
time GDB reaches the symbol main we have only executed one 'nop'
instruction, and the XMM registers should all have the value 0. I've
extended the test script to confirm that $xmm0 to $xmm15 are all
initially 0, and I also check that at the point after $xmm0 has been
modified, all the other XMM registers ($xmm1 to $xmm15) are still 0.
The test program is still linked against libc in order that we can
call the exit function, however, we now call _exit rather than exit in
order to avoid all of the usual cleanup that exit does. This clean up
tries to tear down things that are usually setup during the startup
code, but now this isn't called calling exit will just result in a
crash.
gdb/testsuite/ChangeLog:
* gdb.arch/amd64-disp-step-avx.S: Add '_start' label.
(done): Call '_exit' not 'exit' to avoid atexit handlers.
* gdb.arch/amd64-disp-step-avx.exp: Pass -static, and
-nostartfiles when compiling the test. Confirm that all registers
xmm0 to xmm15 are initially 0, and that xmm1 to xmm15 are 0 after.
A small number of tests incorrectly tried to pass -Wa,-g through to
GCC as an extra compile time flag, either to gdb_compile or
prepare_for_testing.
The problem is that the syntax used for passing the flags was
incorrect, and as a result these extra flags were being ignored.
Luckily, the 'debug' flag was being passed in each case anyway, which
means that the '-g' flag would already be added.
Given that all these tests pass 'debug', and the invalid flag has been
ignored for some time, I'm just removing the flags in this commit.
I've also changed the tests from using gdb_compile to
prepare_for_testing, which allows some extra code to be removed from a
couple of tests scripts.
There should be no change in the test results after this commit.
gdb/testsuite/ChangeLog:
* gdb.arch/amd64-disp-step-avx.exp: Remove unneeded assembler flag
option, syntax was wrong anyway.
* gdb.arch/arm-disp-step.exp: Likewise.
* gdb.arch/sparc64-regs.exp: Likewise.
* gdb.arch/amd64-disp-step.exp: Remove unneeded assembler flag
option, syntax was wrong anyway, switch to use
prepare_for_testing.
* gdb.arch/i386-disp-step.exp: Likewise.
This removes some cleanups from record-full.c in a straightforward
way.
Tested by the buildbot.
gdb/ChangeLog
2018-03-22 Tom Tromey <tom@tromey.com>
* record-full.c (record_full_exec_insn): Use gdb::byte_vector.
(record_full_goto_bookmark): Use std::string.
This patch merges the masks for matching the stwux and stdux
instructions in rs6000-tdep.c:skip_prologue into a single mask that
only matches these two instructions.
Commit 72dd273062 fixed the warning described in PR tdep/18295, this
patch addresses the comment in the same PR indicating that the mask
was too permissive.
gdb/Changelog:
PR tdep/18295
* rs6000-tdep.c (skip_prologue): Match both stwux and stdux
a single mask.
This patch extends rs6000-tdep.c:skip_prologue so that it can detect
when the Link Register is saved using the frame pointer (usually r31)
in adition to the stack pointer (r1).
The frame pointer offset from the frame base is tracked separately
from the stack pointer offset for cases when the frame pointer is not
in synch with the stack pointer at the moment of the LR save.
Previously, "stq" could also be detected as an instruction that saves
LR or CR. Because this was likely unintentional, this patch also
restricts the matches to stw/stwu/std/stdu.
gdb/ChangeLog:
* rs6000-tdep.c (store_insn_p): New function.
(skip_prologue): New variable alloca_reg_offset. Set lr_reg
and cr_reg to their unshifted values. Use store_insn_p to
match LR saves using either R1 or fdata->alloca_reg. Use
store_insn_p to match CR saves. Set alloca_reg_offset
when alloca_reg and framep are set. Remove lr_reg shift
when assigning to fdata->lr_register.
Currently "info proc cmdline" on GNU/Linux does not show the full command
line, but only argument 0. And even a warning is shown if there are more.
This was discussed in 2014 already:
https://sourceware.org/ml/gdb-patches/2014-04/msg00212.html
Follow the advice there and avoid target_fileio_read_stralloc. Instead,
use target_fileio_read_alloc to read the whole command line and then
replace NUL characters by spaces. Also add an appropriate test case.
Note that gdbserver already handles this correctly.
gdb/ChangeLog:
* linux-tdep.c (linux_info_proc): For "info proc cmdline", print
command line args instead of emitting a warning.
gdb/testsuite/ChangeLog:
* gdb.base/info-proc.exp: Add test for "info proc cmdline".
This corrects bad formatting in the newly introduced function
s390_get_wordsize.
gdb/gdbserver/ChangeLog:
* linux-s390-low.c (s390_get_wordsize): Correct brace style.
A few of them were missed in commit 7ac2002247 ("x86: derive DispN
from BaseIndex") and also couldn't be removed by subsequent commits
touching certain templates anyway.
Since they're shorter to encode, the 0xa0...0xa3 encodings are preferred
for moves between accumulator and absolute address outside of 64-bit
mode. With HLE release semantics this encoding is unsupported though,
with the assembler raising an error. The operation is valid though, we
merely need to pick the longer encoding in that case.
The wrong placement of the Load attribute in the templates prevented
this from working. The disassembler also didn't handle this consistently
with other similar dual-encoding insns.
While many templates allowing multiple suitably matching XMM/YMM/ZMM
operand sizes can be folded, a few need to be split in order to not
wrongly accept "xmmword ptr" operands when only XMM registers are
permitted (and memory operands are more narrow). Add a test case
validating this.
This patch replaces VEC(static_tracepoint_marker_p) with std::vector,
and does some c++ification around that. I thought a new overload of
hex2str was useful, so I added it as well as corresponding unit tests.
I also added an overload of ui_out::field_string that takes an
std::string directly.
gdb/ChangeLog:
* tracepoint.h (struct static_tracepoint_marker): Initialize
fields, define default constructor, move constructor and move
assignment, disable the rest.
<str_id, extra>: Make std::string.
(release_static_tracepoint_marker): Remove.
(free_current_marker): Remove.
* tracepoint.c (free_current_marker): Remove.
(parse_static_tracepoint_marker_definition): Adjust to
std::string, use new hex2str overload.
(release_static_tracepoint_marker): Remove.
(print_one_static_tracepoint_marker): Get marker by reference
and adjust to std::string.
(info_static_tracepoint_markers_command): Adjust to std::vector
changes
* target.h (static_tracepoint_marker_p): Remove typedef.
(DEF_VEC_P(static_tracepoint_marker_p)): Remove.
(struct target_ops) <to_static_tracepoint_marker_at>: Return
bool.
<to_static_tracepoint_markers_by_strid>: Return std::vector.
* target-debug.h
(target_debug_print_VEC_static_tracepoint_marker_p_p): Remove.
(target_debug_print_std_vector_static_tracepoint_marker): New.
(target_debug_print_struct_static_tracepoint_marker_p): Rename
to...
(target_debug_print_static_tracepoint_marker_p): ... this.
* target-delegates.c: Re-generate.
* breakpoint.h (struct tracepoint) <static_trace_marker_id>:
Make std::string.
* breakpoint.c (init_breakpoint_sal): Adjust to std::string.
(decode_static_tracepoint_spec): Adjust to std::vector.
(tracepoint_print_one_detail): Adjust to std::string.
(strace_marker_decode_location): Adjust to std::string.
(update_static_tracepoint): Adjust to std::string, remove call
to release_static_tracepoint_marker.
* linux-nat.c (linux_child_static_tracepoint_markers_by_strid):
Adjust to std::vector.
* remote.c (remote_static_tracepoint_marker_at): Return bool.
(remote_static_tracepoint_markers_by_strid): Adjust to
std::vector.
* common/rsp-low.h (hex2str): New overload with explicit count
of bytes.
* common/rsp-low.c (hex2str): New overload with explicit count
of bytes.
* unittests/rsp-low-selftests.c (test_hex2str): New function.
(_initialize_rsp_low_selftests): Add test_hex2str test.
* unittests/tracepoint-selftests.c
(test_parse_static_tracepoint_marker_definition): Adjust to
std::string.
Since I modify the parse_static_tracepoint_marker_definition function in
the next patch, I wanted to write a unit test for it. Doing so showed
that it doesn't handle multiple consecutive static tracepoint
definitions separated by commas. However, the RSP documentation [1]
states that servers may return multiple definitions, like:
1234:6d61726b657231:6578747261207374756666,abba:6d61726b657232:
The problem is that the function uses strlen to compute the length of
the last field (the extra field). If there are additional definitions
in addition to the one we are currently parsing, the returned length
will include those definitions, and we'll try to hex-decode past the
extra field.
This patch changes parse_static_tracepoint_marker_definition to consider
the case where the current definition is followed by a comma and more
definitions. It also adds the unit test that found the issue in the
first place.
I don't think this causes any backwards compatibility issues, because
the previous code only handled single static tracepoint definitions, and
the new code handles that correctly.
gdb/ChangeLog:
* tracepoint.c (parse_static_tracepoint_marker_definition):
Consider case where the definition is followed by more
definitions.
* Makefile.in (SUBDIR_UNITTESTS_SRCS): Add
tracepoint-selftests.c.
* unittests/tracepoint-selftests.c: New.
[1] https://sourceware.org/gdb/onlinedocs/gdb/Tracepoint-Packets.html#qTfSTM
The following bits in DT_FLAGS_1 are used by Solaris:
#define DF_1_KMOD 0x10000000 /* kernel module */
#define DF_1_WEAKFILTER 0x20000000 /* apply weak binding to DT_FILTER */
#define DF_1_NOCOMMON 0x40000000 /* no COMMON symbols exist */
binutils/
* readelf.c (process_dynamic_section): Also dump DF_1_KMOD,
DF_1_WEAKFILTER and DF_1_NOCOMMON.
include/
* elf/common.h (DF_1_KMOD): New.
(DF_1_WEAKFILTER): Likewise.
(DF_1_NOCOMMON): Likewise.
The in-process agent does not handle tdescs with guarded storage yet.
This is fixed.
gdb/gdbserver/ChangeLog:
* linux-s390-ipa.c (get_ipa_tdesc): Add handling for
S390_TDESC_GS.
* linux-s390-low.c (s390_get_ipa_tdesc_idx): Likewise.
(initialize_low_tracepoint): Call init_registers_s390x_gs_linux64
and init_registers_s390_gs_linux64.
