In gdb.trace/unavailable.exp, an action is defined to collect
struct_b.struct_a.array[2] and struct_b.struct_a.array[100],
struct StructB
{
int d, ef;
StructA struct_a;
int s:1;
static StructA static_struct_a;
const char *string;
};
and the other files are not collected.
When GDB examine traceframe collected by the action, "struct_b" is
unavailable completely, which is wrong.
(gdb) p struct_b
$1 = <unavailable>
When GDB reads 'struct_b', it will request to read memory at struct_b's address
of length LEN. Since struct_b.d is not collected, no 'M' block
includes the first part of the desired range, so tfile_xfer_partial returns
TARGET_XFER_UNAVAILABLE and GDB thinks the whole requested range is unavailable.
In order to fix this problem, in the iteration to 'M' blocks, we record the
lowest address of blocks within the request range. If it has, the requested
range isn't unavailable completely. This applies to ctf too. With this patch
applied, the result looks good and fails in unavailable.exp is fixed.
(gdb) p struct_b
$1 = {d = <unavailable>, ef = <unavailable>, struct_a = {a = <unavailable>, b = <unavailable>, array = {<unavailable>,
<unavailable>, -1431655766, <unavailable> <repeats 97 times>, -1431655766, <unavailable> <repeats 9899 times>}, ptr = <unavailable>, bitfield = <unavailable>}, s = <unavailable>, static static_struct_a = {a = <unavailable>, b = <unavailable>, array = {<unavailable> <repeats 10000 times>}, ptr = <unavailable>,
bitfield = <unavailable>}, string = <unavailable>}
gdb:
2014-05-05 Yao Qi <yao@codesourcery.com>
Pedro Alves <palves@redhat.com>
* tracefile-tfile.c (tfile_xfer_partial): Record the lowest
address of blocks that intersects the requested range. Trim
LEN up to LOW_ADDR_AVAILABLE if read from executable read-only
sections.
* ctf.c (ctf_xfer_partial): Likewise.
gdb/testsuite:
2014-05-05 Yao Qi <yao@codesourcery.com>
* gdb.trace/unavailable.exp (gdb_collect_args_test): Save
traceframes into tfile and ctf trace files. Read data from
trace file and test collected data.
(gdb_collect_locals_test): Likewise.
(gdb_unavailable_registers_test): Likewise.
(gdb_unavailable_floats): Likewise.
(gdb_collect_globals_test): Likewise.
(top-level): Append "ctf" to trace_file_targets if GDB
supports.
When I run refactored unavailable.exp, I find
command display behaves a little different on live inferior and on
examining traceframes. In live inferior, when command "display argc"
is typed, the value of "argc" is shown.
(gdb) display argc
1: argc = 1 '\001'
however, on tfile target, when command "display argc" is typed, the
value of "argc" is not shown.
(gdb) tfind
Found trace frame 0, tracepoint 1
at ../../../../git/gdb/testsuite/gdb.trace/unavailable.cc:198
198 i = (int) argc + argi + argf + argd + argstruct.memberi + argarray[1];
(gdb) display argc
I also notice that on "core" target, the value of "argc" isn't shown
either. This difference is caused by the code below in printcmd.c:display_command,
if (from_tty && target_has_execution)
do_one_display (new);
Looks the value of each display is shown if the target has execution.
Source code archaeology doesn't tell much about this requirement.
However, if we type command "display" then on "core" or "tfile"
target, the value of "argc" is still displayed,
for "core" target,
(gdb) display argc
(gdb) display
1: argc = 1 '\001'
for "tfile" target,
(gdb) display argc
(gdb) display
1: argc = <unavailable>
I feel that it is not necessary to have such "target has execution"
requirement to show the value of new created display. Auto-display is
a feature to show the value of expression frequently, has nothing to
do with whether target has execution or not. On the other hand, GDB
has the requirement for new created display, but command "display" can
still show them, this is an inconsistency, which should be fixed.
This patch is to remove the checking to target_has_execution from the
condition.
gdb:
2014-05-05 Yao Qi <yao@codesourcery.com>
* printcmd.c (display_command): Remove the check to
target_has_execution.
This patch moves traceframe checking code out of traceframe generation,
so that we can generation traceframe once, and do the checking in multiple
times (with target remote, tfile and ctf respectively). This is a
pure refactor, not functional changes in unavailable.exp.
gdb/testsuite:
2014-05-05 Yao Qi <yao@codesourcery.com>
* gdb.trace/unavailable.exp (gdb_collect_args_test): Move some
code to ...
(gdb_collect_args_test_1): ... it. New proc.
(gdb_collect_locals_test): Move some code to ...
(gdb_collect_locals_test_1): ... it. New proc.
(gdb_unavailable_registers_test): Move some code to ...
(gdb_unavailable_registers_test_1): ... it. New proc.
(gdb_unavailable_floats): Move some code to ...
(gdb_unavailable_floats_1): ... it. New proc.
This commit is actually an update to make the parser in
gdb/stap-probe.c be aware of all the possible prefixes that a probe
argument can have. According to the section "Argument Format" in:
<https://sourceware.org/systemtap/wiki/UserSpaceProbeImplementation>
The bitness of the arguments can be 8, 16, 32 or 64 bits, signed or
unsigned. Currently GDB recognizes only 32 and 64-bit arguments.
This commit extends this. It also provides a testcase, only for
x86_64 systems.
gdb/
2014-05-02 Sergio Durigan Junior <sergiodj@redhat.com>
* stap-probe.c (enum stap_arg_bitness): New enums to represent 8
and 16-bit signed and unsigned arguments. Update comment.
(stap_parse_probe_arguments): Extend code to handle such
arguments. Use warning instead of complaint to notify about
unrecognized bitness.
gdb/testsuite/
2014-05-02 Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.arch/amd64-stap-optional-prefix.S (main): Add several
probes to test for bitness recognition.
* gdb.arch/amd64-stap-optional-prefix.exp
(test_probe_value_without_reg): New procedure.
Add code to test for different kinds of bitness.
This commit fixes PR breakpoints/16889, which is about a bug that
triggers when GDB tries to parse probes whose arguments do not contain
the initial (and optional) "N@" part. For reference sake, the de
facto format is described here:
<https://sourceware.org/systemtap/wiki/UserSpaceProbeImplementation>
Anyway, this PR actually uncovered two bugs (related) that were
happening while parsing the arguments. The first one was that the
parser *was* catching *some* arguments that were missing the "N@"
part, but it wasn't correctly setting the argument's type. This was
causing a NULL pointer being dereferenced, ouch...
The second bug uncovered was that the parser was not catching all of
the cases for a probe which did not provide the "N@" part. The fix
for that was to simplify the check that the code was making to
identify non-prefixed probes. The code is simpler and easier to read
now.
I am also providing a testcase for this bug, only for x86_64
architectures.
gdb/
2014-05-02 Sergio Durigan Junior <sergiodj@redhat.com>
PR breakpoints/16889
* stap-probe.c (stap_parse_probe_arguments): Simplify
check for non-prefixed probes (i.e., probes whose
arguments do not start with "N@"). Always set the
argument type to a sane value.
gdb/testsuite/
2014-05-02 Sergio Durigan Junior <sergiodj@redhat.com>
PR breakpoints/16889
* gdb.arch/amd64-stap-optional-prefix.S: New file.
* gdb.arch/amd64-stap-optional-prefix.exp: Likewise.
* gas/i386/opcode-intel.d: Undo the last change.
* gas/i386/opcode-suffix.d: Likewise.
* gas/i386/opcode.d: Likewise.
* gas/i386/opcode.s: Likewise.
* gas/i386/prefix.s: Add test for fwait with prefix.
* gas/i386/prefix.d: Updated.
sigsetjmp/siglongjmp without saving the signal mask is faster than
setjmp/longjmp on systems where the signal mask is saved. This patch
uses sigsetjmp/siglongjmp without saving the signal mask if possible.
PR binutils/16886
* config.in: Regenerated.
* configure: Likewise.
* configure.in: Check if sigsetjmp is available.
* h8500-dis.c (private): Replace jmp_buf with OPCODES_SIGJMP_BUF.
(fetch_data): Replace longjmp with OPCODES_SIGLONGJMP.
(print_insn_h8500): Replace setjmp with OPCODES_SIGSETJMP.
* i386-dis.c (dis_private): Replace jmp_buf with OPCODES_SIGJMP_BUF.
(fetch_data): Replace longjmp with OPCODES_SIGLONGJMP.
(print_insn): Replace setjmp with OPCODES_SIGSETJMP.
* ns32k-dis.c (private): Replace jmp_buf with OPCODES_SIGJMP_BUF.
(fetch_data): Replace longjmp with OPCODES_SIGLONGJMP.
(print_insn_ns32k): Replace setjmp with OPCODES_SIGSETJMP.
* sysdep.h (OPCODES_SIGJMP_BUF): New macro.
(OPCODES_SIGSETJMP): Likewise.
(OPCODES_SIGLONGJMP): Likewise.
* vax-dis.c (private): Replace jmp_buf with OPCODES_SIGJMP_BUF.
(fetch_data): Replace longjmp with OPCODES_SIGLONGJMP.
(print_insn_vax): Replace setjmp with OPCODES_SIGSETJMP.
* xtensa-dis.c (dis_private): Replace jmp_buf with
OPCODES_SIGJMP_BUF.
(fetch_data): Replace longjmp with OPCODES_SIGLONGJMP.
(print_insn_xtensa): Replace setjmp with OPCODES_SIGSETJMP.
* z8k-dis.c(instr_data_s): Replace jmp_buf with OPCODES_SIGJMP_BUF.
(fetch_data): Replace longjmp with OPCODES_SIGLONGJMP.
(print_insn_z8k): Replace setjmp with OPCODES_SIGSETJMP.
In a test I was writting, I needed a procedure that would connect to
the target, and do "load", or equivalent.
Years ago, boards would override gdb_load to implement that. Then
gdb_reload was added, and gdb_load was relaxed to allow boards avoid
the spawing and connecting to the target. This sped up gdbserver
testing. See
https://www.sourceware.org/ml/gdb-patches/2007-02/msg00318.html.
To actually spawn the target and load the executable on the target
side, gdb_reload was born:
# gdb_reload -- load a file into the target. Called before "running",
# either the first time or after already starting the program once,
# for remote targets. Most files that override gdb_load should now
# override this instead.
proc gdb_reload { } {
# For the benefit of existing configurations, default to gdb_load.
# Specifying no file defaults to the executable currently being
# debugged.
return [gdb_load ""]
}
Note the comment about specifying no file. Indeed looking at
config/sid.exp, or config/monitor.exp, we see examples of that.
However, the default gdb_load itself doesn't handle the case of no
file specified. When passed no file, it just calls gdb_file_cmd with
no file either, which ends up invocing the "file" command with no
argument, which means unloading the file and its symbols... That
means calling gdb_reload when testing against native targets is
broken. We don't see that today because the only call to gdb_reload
that exists today is guarded by target_info exists
gdb,do_reload_on_run.
The native-extended-gdbserver.exp board is likewise broken here. When
[gdb_load ""] is called, the board sets the remote exec-file to "" ...
Tested on x86_64 Fedora 17, native, remote gdbserver and
extended-remote gdbserver.
testsuite/
2014-05-01 Pedro Alves <palves@redhat.com>
* lib/gdb.exp (gdb_load): Extend comment. Skip calling
gdb_file_cmd if no file is specified.
* boards/native-extended-gdbserver.exp (gdb_load): Use the
last_loaded_file to set the remote exec-file.
of address.
(mmo_write_chunk): When handling data remainder, assert that
previous remaining data is flushed.
(mmo_write_loc_chunk): Only look for trailing and leading zeros
when dealing with an aligned VMA and for aligned lengths. Don't skip
the last 32-bit-word of zeros.
(mmo_write_loc_chunk): Emit an error if the VMA is not aligned.
(mmo_scan) <case LOP_QUOTE>: Move re-alignment of vma before
emitting data, not after updating it.
<case LOP_LOC>: Call mmo_decide_section with aligned vma.
When connecting to a remote system, we use the compare-sections
command to verify that the box is running the code that we think it is
running. Since the system is up and running and *NOT* 'freshly
downloaded without yet executing anything', read-write sections, of
course, differ from what they were in the executable file.
Comparing read-write sections takes time and more importantly the
MIS-MATCHED output is confusing to some users.
The compare-sections command compares all loadable sections including
read-write sections. This patch gives the user the option to compare
just the loadable read-only sections.
gdb/
2014-05-01 David Taylor <dtaylor@emc.com>
* remote.c (compare_sections_command): Add -r option to compare
all loadable read-only sections.
gdb/doc/
2014-05-01 David Taylor <dtaylor@emc.com>
* gdb.texinfo (compare-sections): Document the new -r (read-only)
option.
0x9b (fwait) is both an instruction and an opcode prefix. When 0x9b is
treated as an instruction, we need to handle any prefixes before it.
This patch handles it properly.
gas/testsuite/
PR binutils/16891
* gas/i386/opcode.s: Add test for fwait with prefix.
* gas/i386/opcode-intel.d: Updated.
* gas/i386/opcode-suffix.d: Likewise.
* gas/i386/opcode.d: Likewise.
opcodes/
PR binutils/16891
* i386-dis.c (print_insn): Handle prefixes before fwait.
This adds a variant of local-remote-host-notty.exp that forces
pseudo-tty allocation, so that readline/editing is enabled.
$ ssh localhost gdb -q
(gdb) show editing
Editing of command lines as they are typed is off.
(gdb)
vs:
$ ssh -t localhost gdb -q
(gdb) show editing
Editing of command lines as they are typed is on.
We now get, e.g.:
Running ../../../src/gdb/testsuite/gdb.base/filesym.exp ...
PASS: gdb.base/filesym.exp: complete on "filesy"
PASS: gdb.base/filesym.exp: completion list for "filesym"
PASS: gdb.base/filesym.exp: set breakpoint at filesym
gdb/testsuite/
2014-05-01 Pedro Alves <palves@redhat.com>
* boards/local-remote-host.exp: New file.
When testing with this board, stdin is not a tty, and so
readline/editing is disabled:
$ ssh localhost gdb -q
(gdb) show editing
Editing of command lines as they are typed is off.
(gdb)
Rename the file, to make room for a version of this board that forces a pseudo-tty.
gdb/testsuite/
2014-05-01 Pedro Alves <palves@redhat.com>
* boards/local-remote-host.exp: Rename to ...
* boards/local-remote-host-notty.exp: ... this.
thought that an uncompressed .debug_str section was compressed.
* compress.c (bfd_is_section_compressed): When checking the
.debug_str section, also check the fifth byte in the section is
not part of a string.
* binutils-all/debug_str.s: New test.
* binutils-all/debug_str.d: New test control file.
* binutils-all/compress.exp: Run debug_str test.
This is all we should need to be able to run the eh_frame parts of
bfd_elf_discard_info before bfd_elf_size_dynamic_sections
* elf-eh-frame.c (struct cie.personality): Replace val with sym.
(find_merged_cie): Identify personality functions by (bfd_id,index)
pair when a local sym is used.
Outgoing section for relocations was computed by setting a shared
pointer to which section should be used. For TLS this was overriden to
use .rela.got since they use GOT entries but since the pointer is per
section that whole section was relocated to .rela.got, even non-TLS
relocations.
* elf32-or1k.c: Fix a bug where non-TLS relocations would be forced
into .rela.got if it contained TLS relocations as well.
... and others. The recent patch that fixed several "set remote
foo-packet on/off" commands introduced a regression, observable when
connecting GDB to QEMU. For instance:
(gdb) set debug remote 1
(gdb) tar rem :4444
Remote debugging using :4444
Sending packet: $qSupported:multiprocess+;qRelocInsn+#2a...Ack
Packet received: PacketSize=1000;qXfer:features:read+
Packet qSupported (supported-packets) is supported
Sending packet: $Hgp0.0#ad...Ack
Packet received: OK
Sending packet: $qXfer:features:read:target.xml:0,ffb#79...Ack
Packet received: [...]
Sending packet: $qXfer:features:read:arm-core.xml:0,ffb#08...Ack
Packet received: [...]
!!! -> Sending packet: $QNonStop:0#8c...Ack
Packet received:
Remote refused setting all-stop mode with:
The "QNonStop" feature is associated with the PACKET_QNonStop packet,
with a default of PACKET_DISABLE, so GDB should not be sending the
packet at all.
The patch that introduced the regression decoupled packet_config's
'detect' and 'support' fields, making the former (an auto_boolean)
purely the associated "set remote foo-packet" command's variable. In
the example above, the packet config's 'supported' field does end up
correctly set to PACKET_DISABLE. However, nothing is presently
initializing packet configs that don't actually have a command
associated. Those configs's 'detect' field then ends up set to
AUTO_BOOLEAN_TRUE, simply because that happens to be 0. This forces
GDB to assume the packet is supported, irrespective of what the target
claims it supports, just like if the user had done "set remote
foo-packet on" (this being the associated command, if there was one).
Ideally, all packet configs would have a command associated. While
that isn't true, make sure all packet configs are initialized, even if
no command is associated, and add an assertion that prevents adding
more packets/features without an associated command.
Tested on x86_64 Fedora 17, against pristine gdbserver, and against a
gdbserver with the QNonStop packet/feature disabled with a local hack.
gdb/
2014-04-29 Pedro Alves <palves@redhat.com>
* remote.c (struct packet_config) <detect>: Extend comment.
(add_packet_config_cmd): Don't set the config's detect or support
fields here.
(init_all_packet_configs): Also initialize the config's 'detect'
field.
(reset_all_packet_configs_support): New function.
(remote_open_1): Call reset_all_packet_configs_support instead of
init_all_packet_configs.
(_initialize_remote): Initialize all packet configs. Assert that
all packets have an associated command, except a few known
outliers.
This add a testcases that verifies correct handling of dynamicity
for lower bounds of arrays.
gdb/testsuite/ChangeLog:
* gdb.ada/dyn_arrayidx: New testcase.
Currently, read_subrange_type handles dynamicity only in the case of
the upper bound, and assumes that the lower bound is always static.
That's rooted in the fact that dynamicity was added to support C99
variable-length arrays, where the lower bound is always zero, and
therefore never dynamic. But the lower bound can, in fact, be dynamic
in other languages such as Ada.
Consider for instance the following declaration in Ada...
type Array_Type is array (L .. U) of Natural;
... where L and U are parameters of the function where the declaration
above was made, and whose value are 5 and 10. Currently, the debugger
is able to print the value of the upper bound correctly, but not the
lower bound:
(gdb) ptype array_type
type = array (1 .. 10) of natural
After this patch, the debugger now prints:
(gdb) ptype array_type
type = array (5 .. 10) of natural
gdb/ChangeLog:
* dwarf2read.c (read_subrange_type): Handle dynamic
DW_AT_lower_bound attributes.
Consider the following declaration in Ada...
type Array_Type is array (L .. U) of Natural;
... where L and U are parameters of the function where the declaration
above was made. At the moment, GDB relies on descriptive types in order
to properly decode the array bounds. For instance, if L was 5, and U
was 10, we would see the following:
(gdb) ptype array_type
type = array (5 .. 10) of natural
(gdb) maintenance set ada ignore-descriptive-types
(gdb) ptype array_type
type = array (1 .. 28544912) of natural
This patch enhances ada_discrete_type_{high,low}_bound to resolve
any dynamicity. This is sufficient to fix the case of the upper bound.
For the lower bound, the dwarf2read module does not handle dynamic
lower bounds yet, but once it does, the lower bound should be correctly
handled as well [1].
gdb/ChangeLog:
* ada-lang.c (ada_discrete_type_high_bound): Resolve the type's
dynamic bounds before computing its upper bound.
(ada_discrete_type_low_bound): Same as above with the lower bound.
[1]: The reason why we do not enhance dwarf2read to handle dynamic
lower bounds ahead of this patch is because it unveils some latent
issues such as this one.
This change breaks down the resolve_dynamic_bounds function which
works only on arrays and its index range types into two functions,
one that resolves range types, and one that resolves arrays (using
the new routine to resolve the array's index range type). The
is_dynamic_type and resolve_dynamic_type function are then re-organized
to handle range types as well.
One small change worth mentioning is the fact that, now that range
types are resolved on their own (rather than in the limited context
of array index types), the resolved range types are created from
a copy of the dynamic range type, rather than from scratch (first
parameter of create_range_type). This allows us to preserve as many
original properties in the resolved type as possible (Eg. the type's
name).
This is preparation work that will help better support dynamic range
types for languages that allow the declaration of such types (Eg. Ada).
gdb/ChangeLog:
* dwarf2read.c (is_dynamic_type): Return true for dynamic
range types. Adjust the array handling implementation to
take advantage of this change.
(resolve_dynamic_range): New function, mostly extracted from
resolve_dynamic_bounds.
(resolve_dynamic_array): New function, mostly extracted from
resolve_dynamic_bounds.
(resolve_dynamic_bounds): Delete.
(resolve_dynamic_type): Reimplement. Add handling of
TYPE_CODE_RANGE types.
ada-varobj.c::ada_varobj_describe_simple_array_child only ever gets
called after all GNAT encodings have been applied to (parent_value,
parent_type). So there is no point in redoing it partially by
checking for parallel XA types again.
gdb/ChangeLog:
* ada-varobj.c (ada_varobj_describe_simple_array_child): Remove
handling of parallel ___XA types.
In ada-lang.c::ada_evaluate_subexp, case OP_VAR_VALUE, when noside
is EVAL_AVOID_SIDE_EFFECTS, the first thing we do is set type as
follow:
type = static_unwrap_type (SYMBOL_TYPE (exp->elts[pc + 2].symbol));
Later on in the same block, we make the same call:
return value_zero
(to_static_fixed_type
(static_unwrap_type (SYMBOL_TYPE (exp->elts[pc + 2].symbol))),
not_lval);
This patch removes the second call, since it should result in the same
type being returned, so no point in making that call again.
gdb/ChangeLog:
* ada-lang.c (ada_evaluate_subexp) <OP_VAR_VALUE>: Remove
unnecessary second call to static_unwrap_type.
The problem was that references to weak function symbols were being
incorrectly biased by definition's offset.
PR gas/16858
* config/tc-i386.c (md_apply_fix): Do not adjust value of
pc-relative fixes against weak symbols.
The makefile rule i386-avx512.c is to generate i386-avx512.c, but it
is written to i386-avx.c by mistake. This patch is to fix this typo.
gdb/gdbserver:
2014-04-28 Yao Qi <yao@codesourcery.com>
* Makefile.in (i386-avx512.c): Fix the typo of generated file
name.
When GDB debug DUMMY_FRAME, SIGTRAMP_FRAME and ARCH_FRAME, even if
"set disassemble-next-line on", it will not output the asm code:
(gdb) set disassemble-next-line on
(gdb) si
<signal handler called>
(gdb)
<signal handler called>
(gdb)
<signal handler called>
So make this patch make they can work together, it will become:
(gdb) si
<signal handler called>
=> 0xffffffff816bfb09 <int_with_check+0>: 65 48 8b 0c 25 c8 c7 00 00 mov %gs:0xc7c8,%rcx
(gdb)
<signal handler called>
=> 0xffffffff816bfb12 <int_with_check+9>: 48 81 e9 d8 1f 00 00 sub $0x1fd8,%rcx
(gdb)
<signal handler called>
=> 0xffffffff816bfb19 <int_with_check+16>: 8b 51 10 mov 0x10(%rcx),%edx
2014-04-27 Hui Zhu <hui@codesourcery.com>
* stack.c (print_frame_info): Call do_gdb_disassembly with
DUMMY_FRAME, SIGTRAMP_FRAME and ARCH_FRAME.
In Dwarf::assemble in dwz.exp, 10 is hard-coded in it,
subprogram {
{name main}
{low_pc main addr}
{high_pc "main + 10" addr}
}
however, the length of main function varies on architectures. The
hard-coded 10 here causes dwz.exp fails on some targets, such as
nios2.
This patch is to add some code to compute the length of function main,
which is similar to what we are doing in entry-values.exp.
gdb/testsuite:
2014-04-26 Yao Qi <yao@codesourcery.com>
* gdb.dwarf2/dwz.exp: Compile main.c to object. Restart GDB
and compute the length of function main. Save it in
$main_length.
(Dwarf::assemble): Use $main_length instead of hard-coded 10.
(top-level): Use gdb_compile to compile objects into
executable and restart GDB. Remove invocation to
prepare_for_testing.
