2015-05-08 Sandra Loosemore <sandra@codesourcery.com>
gdb/gdbserver/
* linux-nios2-low.c: Include elf/common.h. Adjust comments.
Remove HAVE_PTRACE_GETREGS conditionals.
(nios2_regsets): Use PTRACE_GETREGSET and PTRACE_SETREGSET
instead of PTRACE_GETREGS and PTRACE_SETREGS.
2015-05-08 Sandra Loosemore <sandra@codesourcery.com>
gdb/
* nios2-tdep.c (nios2_breakpoint_from_pc): Revert to using
"trap 31" as the breakpoint instruction on all targets.
In my last commit to make gdb.base/coredump-filter.exp be more robust
regarding using arrays in the global namespace, I cleared the
"coredump_var_addr" array like this:
set coredump_var_addr ""
# use coredump_var_addr as an array...
This causes DejaGNU to complain because the variable is first set as
non-array, and the used as an array. The correct way to do this is to
unset the variable using:
unset -nocomplain coredump_var_addr
# use coredump_var_addr as an array...
The "-nocomplain" part is necessary because if the variable doesn't
exist "unset" will not error.
Tested on Fedora 20 x86_64.
gdb/testsuite/ChangeLog:
2015-05-08 Sergio Durigan Junior <sergiodj@redhat.com>
* gdb.base/coredump-filter.exp: Correctly unset
"coredump_var_addr" array.
Attempting to build libiberty on LynxOS-178 fails trying to compile
mkstemps.c with the following error:
mkstemps.c:84:18: error: storage size of 'tv' isn't known
struct timeval tv;
^
This file would normally include <sys/time.h> to get the type's
definition, but unfortunately LynxOS-178 does not want us to use
<sys/time.h>, only <time.h>. The configure script correctly finds
this out and generates a config.h file where HAVE_SYS_TIME_H is
undefined:
/* Define to 1 if you have the <sys/time.h> header file. */
/* #undef HAVE_SYS_TIME_H */
This patch fixes the build issue by falling back on including <time.h>
if <sys/time.h> could not be included (and provided that HAVE_TIME_H
is defined, of course).
libiberty/ChangeLog:
* mkstemps.c: #include <time.h> if HAVE_TIME_H is defined
but not HAVE_SYS_TIME_H.
Sequential test runs are stopping prematurely like this:
$ make check RUNTESTFLAGS="non-existing-program.exp server-exec-info.exp"
Running /home/pedro/gdb/mygit/src/gdb/testsuite/gdb.server/non-existing-program.exp ...
Running /home/pedro/gdb/mygit/src/gdb/testsuite/gdb.server/server-exec-info.exp ...
can not find channel named "exp6"
while executing
"match_max [match_max -d]"
(procedure "default_gdb_init" line 26)
invoked from within
"default_gdb_init $test_file_name"
(procedure "gdb_init" line 83)
invoked from within
"${tool}_init $test_file_name"
(procedure "runtest" line 18)
invoked from within
"runtest $test_name"
("foreach" body line 42)
invoked from within
...
make[2]: *** [check-single] Error 1
make[2]: Leaving directory `/home/pedro/gdb/mygit/build/gdb/testsuite'
make[1]: *** [check] Error 2
make[1]: Leaving directory `/home/pedro/gdb/mygit/build/gdb/testsuite'
make: *** [check] Error 2
default_gdb_init has this:
# Unlike most tests, we have a small number of tests that generate
# a very large amount of output. We therefore increase the expect
# buffer size to be able to contain the entire test output. This
# is especially needed by gdb.base/info-macros.exp.
match_max -d 65536
# Also set this value for the currently running GDB.
match_max [match_max -d]
It's the second match_max that is erroring. As that call does not
specify an explicit channel name with -i, expect defaults to
$spawn_id, which is pointing at a channel that is already gone. (If
the spawn_id variable is not set, match_max defaults to
$user_spawn_id / stdin/out).
gdb/testsuite/ChangeLog:
2015-05-08 Pedro Alves <palves@redhat.com>
* gdb.server/non-existing-program.exp: Unset spawn_id.
PR gas/18347
* config/tc-arm.h (TC_EQUAL_IN_INSN): Define.
* config/tc-arm.c (arm_tc_equal_in_insn): New function. Move
the symbol name checking code to here from...
(md_undefined_symbo): ... here.
Consider the following declarations...
type Obj_T (Selected_Flights_Length : Natural) is
record
Selected_Flights : Flights.List.T (1 .. Selected_Flights_Length);
end record;
Broken : Obj_T;
... which defines a variable named "broken" which is a discrimated
record where broken.Selected_Flights is an array whose upper bound
is stored in the record's Selected_Flights_Length component.
Trying to print the value of that object currently fails:
(gdb) print broken
cannot find reference address for offset property
Looking at the debugging info, we see that variable "Broken" is
a reference...
<1><8e3>: Abbrev Number: 21 (DW_TAG_const_type)
<8e4> DW_AT_type : <0x8e8>
<1><8e8>: Abbrev Number: 22 (DW_TAG_reference_type)
<8e9> DW_AT_byte_size : 8
<8ea> DW_AT_type : <0x7ec>
... to ...
<1><7ec>: Abbrev Number: 12 (DW_TAG_structure_type)
<7ed> DW_AT_name : (indirect string, offset: 0xc6d): reprod__obj_t
<7f1> DW_AT_decl_file : 2
<7f2> DW_AT_decl_line : 15
<7f3> DW_AT_GNAT_descriptive_type: <0x87e>
<7f7> DW_AT_sibling : <0x87e>
... which has 2 members, the first one being the discriminant...
<2><7fb>: Abbrev Number: 9 (DW_TAG_member)
<7fc> DW_AT_name : (indirect string, offset: 0xc98): selected_flights_length
<800> DW_AT_decl_file : 2
<801> DW_AT_decl_line : 15
<802> DW_AT_type : <0x807>
<806> DW_AT_data_member_location: 0
... and the second one being the one that causes trouble...
<2><83d>: Abbrev Number: 9 (DW_TAG_member)
<83e> DW_AT_name : (indirect string, offset: 0xd17): selected_flights
<842> DW_AT_decl_file : 2
<843> DW_AT_decl_line : 17
<844> DW_AT_type : <0x815>
<848> DW_AT_data_member_location: 4
The second field's type is an array....
<2><815>: Abbrev Number: 14 (DW_TAG_array_type)
<816> DW_AT_name : (indirect string, offset: 0xd2f): reprod__obj_t__T5sP
<81a> DW_AT_GNAT_descriptive_type: <0x7e1>
<81e> DW_AT_type : <0x748>
<822> DW_AT_sibling : <0x830>
... whose uppper bound is a reference to <0x7fb>...
<3><826>: Abbrev Number: 15 (DW_TAG_subrange_type)
<827> DW_AT_type : <0x830>
<82b> DW_AT_upper_bound : <0x7fb>
<3><82f>: Abbrev Number: 0
Because the upper bound is dynamic, we try to resolve it.
As it happens, the upper-bound resolution for this range type
works fine. What breaks is when we try to resolve this range
type's target type, which is:
<2><830>: Abbrev Number: 16 (DW_TAG_subrange_type)
<831> DW_AT_upper_bound : <0x7fb>
<835> DW_AT_name : (indirect string, offset: 0xc7b): reprod__obj_t__T4s___XDLU_1__selected_flights_length
<839> DW_AT_type : <0x766>
<83d> DW_AT_artificial : 1
It is actually pretty much the same as the first subrange type,
so you might ask why this is causing trouble, when the resolution
of the previous DIE worked like a charm???
Well, for that, we need to backtrack a bit, and notice that, ahead
of the DW_TAG_structure_type's DIE, there is the following DIE:
<1><7e1>: Abbrev Number: 6 (DW_TAG_typedef)
<7e2> DW_AT_name : (indirect string, offset: 0xbae): reprod__obj_t__T5s
<7e6> DW_AT_decl_file : 2
<7e7> DW_AT_decl_line : 17
<7e8> DW_AT_type : <0x849>
... and that DIE references an array type...
<2><849>: Abbrev Number: 14 (DW_TAG_array_type)
<84a> DW_AT_name : (indirect string, offset: 0xbae): reprod__obj_t__T5s
<84e> DW_AT_GNAT_descriptive_type: <0x864>
<852> DW_AT_type : <0x748>
<856> DW_AT_sibling : <0x864>
... whose subrange is...
<3><85a>: Abbrev Number: 15 (DW_TAG_subrange_type)
<85b> DW_AT_type : <0x830>
<85f> DW_AT_upper_bound : <0x7fb>
... where the subrange's base type is the DW_TAG_subrange_type DIE
that is causing problem.
In summary, we process the typedef first, which causes us to process
the second subrange BEFORE we process the struct DIE itself, and
therefore the struct's discriminent (DW_TAG_member #1). As a result,
while trying to handle the reference to that DW_TAG_member #1 as
the upper bound of the second range type, we do...
case DW_AT_data_member_location:
{
[...]
baton->referenced_type = get_die_type (target_die->parent,
target_cu);
... where target_die->parent (DW_TAG_member #1) hasn't been processed
yet, and thus get_die_type returns NULL.
This is what later causes us problems trying to find the right address
to use as the base address for our field, which then triggers the
error message we are seeing.
This patch fixes the issue by calling read_type_die instead of
get_die_type. If the DIE has already been processed, then this
is the same as get_die_type. If not, the it'll get the parent
die to be read, and then get its type.
gdb/ChangeLog:
* dwarf2read.c (attr_to_dynamic_prop)
<DW_AT_data_member_location>: Use read_type_die isntead of
get_die_type.
Tested on x86_64-linux, no regression.
No testcase, unfortunately, as the reproducer was given to us by
a customer, and it's been otherwise surprisingly difficult to
reproduce the same error outside of that reproducer.
We observed on x86-windows that trying to call a function from
GDB leads to a mysterious "Invalid cast" error. This can be
observed in gdb.ada/float_param.exp:
(gdb) call set_long_double(1, global_small_struct, 4.0)
Invalid cast.
This happens because the 3rd parameter, a Long_Long_Float, is
actually passed wrapped inside a PAD structure. As documented
in GNAT's exp_dbug.ads, PAD types are simple wrappers that GNAT
uses to handle types with size or alignment constraints.
We already support those when printing an object encapsulated
in a PAD type, but not when trying to pass an argument that
is wrapped inside a PAD type. As a result, what happens is that
call_function_by_hand ends up with an argument with a type
that looks incompatible with the expected type of the argument.
The error comes when trying to push the arguments in inferior
memory, while trying to coerce each one of them to their expected
types (in value_arg_coerce).
Note that the problem is not specific to Windows, but so far, this is
the only platform where we've seen this happen.
gdb/ChangeLog:
* ada-lang.c (ada_convert_actual): Add handling of formals
passed inside an aligner type.
Tested on x86-windows (AdaCore testsuite) and x86_64-linux (official
testsuite as well as AdaCore's testsuite).
Now that the erc32 files have been updated to contain an FSF copyright
header, these files should no longer be in the exclude list.
gdb/ChangeLog:
* copyright.py (NOT_FSF_LIST): Remove sim/erc32 entries.
On ELF target, the assembler normally generates code which can go into a
shared library where non-weak symbols can be preempted. The -mno-shared
option tells the assembler to generate code not for a shared library,
where non-weak symbols won't be preempted. The resulting code is slightly
smaller. This option mainly affects the handling of branch instructions.
gas/
* config/tc-i386.c (no_shared): New.
(OPTION_MNO_SHARED): Likewise.
(elf_symbol_resolved_in_segment_p): Check no_shared.
(md_longopts): Add mno-shared.
(md_parse_option): Handle OPTION_MNO_SHARED.
(md_show_usage): Add -mno-shared.
* doc/c-i386.texi: Document -mno-shared.
gas/testsuite/
* gas/i386/i386.exp: Run relax-4 and x86-64-relax-3.
* gas/i386/relax-4.d: New file.
* gas/i386/x86-64-relax-3.d: Likewise.
Hi,
We see some fails in gdb.base/coredump-filter.exp when we do remote
gdbserver testing, like what I did for arm/aarch64 linux testing or
run it with board file remote-gdbserver-on-localhost
$ make check RUNTESTFLAGS='--target_board=remote-gdbserver-on-localhost coredump-filter.exp'
we find that this line in the test doesn't work as expected,
remote_exec target "sh -c \"echo $filter_flag > /proc/$ipid/coredump_filter\""
although such pattern has been used in gdb testsuite somewhere else,
but the special thing here is that we redirect the output to
/proc/$ipid/coredump_filter on the remote target. DejaGNU will
redirect the output from the remote target to local, and looks tcl
gets confused by these two redirection.
After trying pass different parameters to remote_exec and hacking
remote_exec/rsh_exec/local_exec, I got no success, I decide
to give up, and try to update /proc/$ipid/coredump_filter by the c
code directly.
This patch adds a c function set_coredump_filter to update
coredump_filter, and GDB calls it.
gdb/testsuite:
2015-05-08 Yao Qi <yao.qi@linaro.org>
PR gdb/18208
* gdb.base/coredump-filter.c (set_coredump_filter): New function.
* gdb.base/coredump-filter.exp (do_save_core): Call inferior
function set_coredump_filter, and remove remote_exec call.
Remove argument ipid. Callers update.
(top level): Don't get inferior's PID.
GDBserver steps over breakpoint if the condition is false, but if target
doesn't support hardware single step, the step over is very simple, if
not incorrect, in linux-arm-low.c:
/* We only place breakpoints in empty marker functions, and thread locking
is outside of the function. So rather than importing software single-step,
we can just run until exit. */
static CORE_ADDR
arm_reinsert_addr (void)
{
struct regcache *regcache = get_thread_regcache (current_thread, 1);
unsigned long pc;
collect_register_by_name (regcache, "lr", &pc);
return pc;
}
and linux-mips-low.c does the same. GDBserver sets a breakpoint at the
return address of the current function, resume and wait the program hits
the breakpoint in order to achieve "breakpoint step over". What if
program hits other user breakponits during this "step over"?
It is worse if the arm/thumb interworking is considered. Nowadays,
GDBserver arm backend unconditionally inserts arm breakpoint,
/* Define an ARM-mode breakpoint; we only set breakpoints in the C
library, which is most likely to be ARM. If the kernel supports
clone events, we will never insert a breakpoint, so even a Thumb
C library will work; so will mixing EABI/non-EABI gdbserver and
application. */
(const unsigned char *) &arm_breakpoint,
(const unsigned char *) &arm_eabi_breakpoint,
note that the comments are no longer valid as C library can be compiled
in thumb mode.
When GDBserver steps over a breakpoint in arm mode function, which
returns to thumb mode, GDBserver will insert arm mode breakpoint by
mistake and the program will crash. GDBserver alone is unable to
determine the arm/thumb mode given a PC address. See how GDB does
it in arm-tdep.c:arm_pc_is_thumb.
After thinking about how to teach GDBserver inserting right breakpoint
(arm or thumb) for a while, I reconsider it from a different direction
that it may be unreasonable to run target-side conditional breakpoint for
targets without hardware single step. Pedro also pointed this out here
https://sourceware.org/ml/gdb-patches/2015-04/msg00337.html
This patch is to add a new target_ops hook
supports_conditional_breakpoints, and only reply
";ConditionalBreakpoints+" if it is true. On linux targets,
supports_conditional_breakpoints returns true if target has hardware
single step, on other targets, (win32, lynx, nto, spu), set it to NULL,
because conditional breakpoint is a linux-specific feature.
gdb/gdbserver:
2015-05-08 Yao Qi <yao.qi@linaro.org>
* linux-low.c (linux_supports_conditional_breakpoints): New
function.
(linux_target_ops): Install new target method.
* lynx-low.c (lynx_target_ops): Install NULL hook for
supports_conditional_breakpoints.
* nto-low.c (nto_target_ops): Likewise.
* spu-low.c (spu_target_ops): Likewise.
* win32-low.c (win32_target_ops): Likewise.
* server.c (handle_query): Check
target_supports_conditional_breakpoints.
* target.h (struct target_ops) <supports_conditional_breakpoints>:
New field.
(target_supports_conditional_breakpoints): New macro.
On 64-bit S390 platforms, for programs compiled with -m31, it could
happen that GDB inadvertently cleared the inferior's 31-bit addressing
mode bit and left the inferior running in 24-bit addressing mode. In
particular this occurred with checkpoint.exp, when the "restore"
command needed to create a new regcache copy: At the time when the
PSWM register was copied over, the addressing mode bit was taken from
the PSWA register, which was still zero since it had not been copied
yet. And when the PSWA register was copied, the addressing mode was
not updated again.
The fix affects fill_gregset, where the bits "belonging" to each of
the PSWA and PSWM registers are now carefully separated. The
addressing mode bit is no longer touched when writing PSWM, and --
more importantly -- it *is* written when writing PSWA.
gdb/ChangeLog:
* s390-linux-nat.c (fill_gregset): Avoid relying on the PSWA
register in the regcache when treating the PSWM register, and vice
versa.
Since watch_thread_num.exp was changed to use access watchpoints, the
test case fails on s390 and s390x, since those targets do not support
access watchpoints. This patch skips the test case on such targets.
gdb/testsuite/ChangeLog:
* gdb.base/watch_thread_num.exp: Skip test on targets without
access watchpoints.
Branches to global non-weak symbols defined in the same segment with
non-default visibility can be optimized the same way as branches to
local symbols.
gas/
* config/tc-i386.c (elf_symbol_resolved_in_segment_p): New.
(md_estimate_size_before_relax): Use it.
gas/testsuite/
* gas/i386/i386.exp: Run relax-3 and x86-64-relax-2.
* gas/i386/relax-3.d: New file.
* gas/i386/relax-3.s: Likewise.
* gas/i386/x86-64-relax-2.d: Likewise.
linux-thread-db.c initializes td_ta_map_id2thr but never uses it.
This commit removes this dead code.
gdb/ChangeLog:
* linux-thread-db.c (struct thread_db_info)
<td_ta_map_id2thr_p>: Remove field.
(try_thread_db_load_1): Remove initialization for the above.
linux-thread-db.c initializes td_thr_validate but never uses it.
This commit removes this dead code.
gdb/ChangeLog:
* linux-thread-db.c (struct thread_db_info)
<td_thr_validate_p>: Remove field.
(try_thread_db_load_1): Remove initialization for the above.
Calling memcpy() could fail as memcpy() from libc is GNU-IFUNC.
gdb/ChangeLog
2015-05-06 Jan Kratochvil <jan.kratochvil@redhat.com>
* compile/compile-object-load.c (compile_object_load): Support
mst_text_gnu_ifunc.
gdb/ChangeLog
2015-05-06 Jan Kratochvil <jan.kratochvil@redhat.com>
* compile/compile.c (compile_to_object): Make the cmd_string parameter
const. Use new variables for the const compatibility.
(eval_compile_command): Make the cmd_string parameter const.
* compile/compile.h (eval_compile_command): Make the cmd_string
parameter const.
$ ./gdbserver :1234 blah
Process blah created; pid = 16471
Cannot exec blah: No such file or directory.
Child exited with status 127
Killing process(es): 16471
../../../../src/binutils-gdb/gdb/gdbserver/linux-low.c:920: A problem internal to GDBserver has been detected.
kill_wait_lwp: Assertion `res > 0' failed.
GDBserver shouldn't even be trying to kill that process. GDBserver
kills or detaches from all processes on exit, and due to a missing
mourn_inferior call, GDBserver tries to kill the process that it had
already seen exit.
Tested on x86_64 Fedora 20. New test included. I emulated what
Windows outputs by hacking an error call in linux_create_inferior.
gdb/gdbserver/ChangeLog:
2015-05-06 Pedro Alves <palves@redhat.com>
PR server/18081
* server.c (start_inferior): If the process exits, mourn it.
gdb/testsuite/ChangeLog:
2015-05-06 Pedro Alves <palves@redhat.com>
PR server/18081
* gdb.server/non-existing-program.exp: New file.
This hook is no longer used, and can therefore be eliminated.
gdb/ChangeLog:
* defs.h (deprecated_init_ui_hook): Delete. Remove associated
comment.
* top.c (deprecated_init_ui_hook): Delete.
(gdb_init): Remove handling of deprecated_init_ui_hook.
* interps.c (clear_interpreter_hooks): Remove handling of
deprecated_init_ui_hook.
* main.c (captured_main): Update comment.
The "info dll", an alias of the "info sharedlibrary" command, is
currently only defined in windows native versions. This patch makes
it universally available by moving the alias declaration to solib.c,
and adjusts the documentation accordingly.
Making it universally available has two benefits:
- Windows users moving to a Unix platforms are still able to use
the same command for getting the list of shared libraries;
- Unix to Windows cross debuggers now provide that command also.
gdb/ChangeLog:
* solib.c (_initialize_solib): Add "info dll" alias creation.
* windows-nat.c (set_windows_aliases): Delete.
(_initialize_windows_nat): Remove deprecated_init_ui_hook
assignment.
* NEWS: Add news entry about "info dll" now being available
on all platforms.
gdb/doc/ChangeLog:
* gdb.texinfo (Files): Add "info dll" documentation.
(Cygwin Native): Remove "info dll" documentation.
gas/ChangeLog:
2015-05-06 Jose E. Marchesi <jose.marchesi@oracle.com>
* config/tc-sparc.c (sparc_ip): Support the %ncc "natural"
condition codes
* doc/c-sparc.texi (Sparc-Regs): Document %ncc.
gas/testsuite/ChangeLog:
2015-05-06 Jose E. Marchesi <jose.marchesi@oracle.com>
* gas/sparc/natural.s: New file.
* gas/sparc/natural-32.s: Likewise.
* gas/sparc/natural.d: Likewise.
* gas/sparc/natural-32.d: Likewise.
* gas/sparc/sparc.exp (sparc_elf_setup): Run the tests natural and
natural-32.
This patch improves the documentation of ada-lang.c's
value_assign_to_component to publish the fact that it also works
with not_lval values.
And touching this area of the code showed that there were a number
of whitespace issues, as well as a formatting issue of the main comment
(no leading '*' on each line). This patch fixes those while at it.
No functional change, however.
gdb/ChangeLog:
* ada-lang.c (value_assign_to_component): Reformat and improve
documentation. Remove all trailing spaces.
This patch improves the handling of out-of-line functions nested
inside functions that have been inlined.
Consider for instance a situation where function Foo_O224_021
has a function Child1 declared in it, which itself has a function
Child2 nested inside Child1. After compiling the program with
optimization on, Child1 gets inlined, but not Child2.
After inserting a breakpoint on Child2, and running the program
until reaching that breakpoint, we get the following backtrace:
% gdb foo_o224_021
(gdb) break foo_o224_021.child1.child2
(gdb) run
[...]
Breakpoint 1, foo_o224_021 () at foo_o224_021.adb:28
28 Child1;
(gdb) bt
#0 0x0000000000402400 in foo_o224_021 () at foo_o224_021.adb:28
#1 0x00000000004027a4 in foo_o224_021.child1 () at foo_o224_021.adb:23
#2 0x00000000004027a4 in foo_o224_021 () at foo_o224_021.adb:28
GDB reports the wrong function name for frame #0. We also get the same
kind of error in the "Breakpoint 1, foo_o224_021 () [...]" message.
In both cases, the function name should be foo_o224_021.child1.child2,
and the parameters should be "s=...".
What happens is that the inlined frame handling does not handle well
the case where an inlined function is calling an out-of-line function
which was declared inside the inlined function's scope.
In particular, looking first at the inlined-frame sniffer when applying
to frame #0:
/* Calculate DEPTH, the number of inlined functions at this
location. */
depth = 0;
cur_block = frame_block;
while (BLOCK_SUPERBLOCK (cur_block))
{
if (block_inlined_p (cur_block))
depth++;
cur_block = BLOCK_SUPERBLOCK (cur_block);
}
What happens is that cur_block starts as the block associated
to child2, which is not inlined. We shoud be stopping here, but
instead, we keep walking the superblock chain, which takes us
all the way to Foo_O224_021's block, via Child2's block. And
since Child1 was inlined, we end up with a depth count of 1,
wrongly making GDB think that frame #0 is an inlined frame.
Same kind of issue inside skip_inline_frames.
The fix is to stop checking for inlined frames as soon as we see
a block corresponding to a function which is not inlined. This is
the behavior we now obtain:
(gdb) run
[...]
Breakpoint 1, foo_o224_021.child1.child2 (s=...) at foo_o224_021.adb:9
9 function Child2 (S : String) return Boolean is
(gdb) bt
#0 0x0000000000402400 in foo_o224_021.child1.child2 (s=...)
at foo_o224_021.adb:9
#1 0x00000000004027a4 in foo_o224_021.child1 () at foo_o224_021.adb:23
#2 0x00000000004027a4 in foo_o224_021 () at foo_o224_021.adb:28
gdb/ChangeLog:
* inline-frame.c (inline_frame_sniffer, skip_inline_frames):
Stop counting inlined frames as soon as an out-of-line function
is found.
gdb/testsuite/ChangeLog:
* gdb.ada/out_of_line_in_inlined.exp: Add run and "bt" tests.
Consider the following code, which defines a function, Child2,
which is itself nested inside Child1:
procedure Foo_O224_021 is
O1 : constant Object_Type := Get_Str ("Foo");
procedure Child1 is
O2 : constant Object_Type := Get_Str ("Foo");
function Child2 (S : String) return Boolean is -- STOP
begin
for C of S loop
Do_Nothing (C);
if C = 'o' then
return True;
end if;
end loop;
return False;
end Child2;
R : Boolean;
begin
R := Child2 ("Foo");
R := Child2 ("Bar");
R := Child2 ("Foobar");
end Child1;
begin
Child1;
end Foo_O224_021;
On x86_64-linux, when compiled at -O2, GDB is unable to insert
a breakpoint on Child2:
% gnatmake -g -O2 foo_o224_021
% gdb foo_o224_021
(gdb) b child2
Function "child2" not defined.
(gdb) b foo_o224_021.child1.child2
Function "foo_o224_021.child1.child2" not defined.
The problem is caused by the fact that GDB did not create a symbol
for Child2, and this, in turn, is caused by the fact that the compiler
decided to inline Child1, but not Child2. The DWARF debugging info
first provides an abstract instance tree for Child1...
<3><1b7b>: Abbrev Number: 29 (DW_TAG_subprogram)
<1b7c> DW_AT_name : (indirect string, offset: 0x23f8): foo_o224_021__child1
<1b82> DW_AT_inline : 1 (inlined)
<1b83> DW_AT_sibling : <0x1c01>
... where that subprogram is given the DW_AT_inline attribute.
Inside that function there is a lexical block which has no PC
range (corresponding to the fact that this is the abstract tree):
<4><1b87>: Abbrev Number: 30 (DW_TAG_lexical_block)
... inside which our subprogram Child2 is described:
<5><1b92>: Abbrev Number: 32 (DW_TAG_subprogram)
<1b93> DW_AT_name : (indirect string, offset: 0x2452): foo_o224_021__child1__child2
<1b99> DW_AT_type : <0x1ab1>
<1b9d> DW_AT_low_pc : 0x402300
<1ba5> DW_AT_high_pc : 0x57
[...]
Then, later on, we get the concrete instance tree, starting at:
<3><1c5e>: Abbrev Number: 41 (DW_TAG_inlined_subroutine)
<1c5f> DW_AT_abstract_origin: <0x1b7b>
<1c63> DW_AT_entry_pc : 0x4025fd
<1c6b> DW_AT_ranges : 0x150
... which refers to Child1. One of that inlined subroutine children
is the concrete instance of the empty lexical block we saw above
(in the abstract instance tree), which gives the actual address
range for this inlined instance:
<5><1c7a>: Abbrev Number: 43 (DW_TAG_lexical_block)
<1c7b> DW_AT_abstract_origin: <0x1b87>
<1c7f> DW_AT_ranges : 0x180
This is the DIE which provides the context inside which we can
record Child2. But unfortunately, GDB does not take the abstract
origin into account when handling lexical blocks, causing it
to miss the fact that this block contains some symbols described
in the abstract instance tree. This is the first half of this patch:
modifying GDB to follow DW_AT_abstract_origin attributes for
lexical blocks.
But this not enough to fix the issue, as we're still unable to
break on Child2 with just that change. The second issue can be
traced to the way inherit_abstract_dies determines the list of
DIEs to inherit from. For that, it iterates over all the DIEs in
the concrete instance tree, and finds the list of DIEs from the
abstract instance tree that are not referenced from the concrete
instance tree. As it happens, there is one type of DIE in the
concrete instance tree which does reference Child2's DIE, but
in fact does otherwise define a concrete instance of the reference
DIE; that's (where <0x1b92> is Child2's DIE):
<6><1d3c>: Abbrev Number: 35 (DW_TAG_GNU_call_site)
<1d3d> DW_AT_low_pc : 0x4026a4
<1d45> DW_AT_abstract_origin: <0x1b92>
So, the second part of the patch is to modify inherit_abstract_dies
to ignore DW_TAG_GNU_call_site DIEs when iterating over the concrete
instance tree.
This patch also includes a testcase which can be used to reproduce
the issue. Unfortunately, for it to actually pass, a smal patch in
GCC is also necessary to make sure that GCC provides lexical blocks'
DW_AT_abstract_origin attributes from the concrete tree back to
the abstract tree. We hope to be able to submit and integrate that
patch in the GCC tree soon. Meanwhile, a setup_xfail has been added.
gdb/ChangeLog:
2014-05-05 Pierre-Marie de Rodat <derodat@adacore.com>
* dwarf2read.c (inherit_abstract_dies): Skip
DW_TAG_GNU_call_site dies while inheriting children of an
abstract DIE into a scope.
(read_lexical_block_scope): Inherit abstract DIE's for
lexical scopes.
gdb/testsuite/ChangeLog:
* gdb.ada/out_of_line_in_inlined: New testcase.
This is an issue which I noticed while working on trying to print
an array of variant records. For instance, trying to print "A1",
an array of elements whose size is variable, defined as follow
(see gdb.ada/var_rec_arr testcase):
subtype Small_Type is Integer range 0 .. 10;
type Record_Type (I : Small_Type := 0) is record
S : String (1 .. I);
end record;
function Ident (R : Record_Type) return Record_Type;
type Array_Type is array (Integer range <>) of Record_Type;
A1 : Array_Type := (1 => (I => 0, S => <>),
2 => (I => 1, S => "A"),
3 => (I => 2, S => "AB"));
The debugger sometimes prints the array as follow:
(gdb) print A1
$1 = ((i => 0, s => ""), (i => 0, s => ""), (i => 0, s => ""))
The problem happens inside the part of the loop printing the array's
elements, while trying to count the number of consecutive elements
that have the same value (in order to replace them by the "<repeats
nnn times>" message when the number exceeds a threshold). In particular,
in ada-valprint.c::val_print_packed_array_elements:
elttype = TYPE_TARGET_TYPE (type);
eltlen = TYPE_LENGTH (check_typedef (elttype));
while (...)
{
if (!value_contents_eq (v0, value_embedded_offset (v0),
v1, value_embedded_offset (v1),
eltlen))
break;
The value comparison is performed using value_contents_eq but makes
the assumption that elttype is not dynamic, which is not always true.
In particular, in the case above, elttype is dynamic and therefore
its TYPE_LENGTH changes from element to element.
As it happens in this case, the eltlen is zero, which causes the call
to value_contents_eq to return true, and therefore GDB thinks all
3 elements of the array are equal.
This patch fixes the issue by making sure that both v0 and v1, which
are values whose type we expect to be resolved, have identical lengths.
If not, then the two elements of the array cannot possibly have the
same value and we do not even need to do the binary comparison.
Unfortunately, this is still not enough to get GDB to print the correct
value for our array, because the assumption that v0 and v1 have a type
which has been resolved is actually not met. So, the second part of
the patch modifies the function that constructed the values to make
sure dynamic types do get resolved.
gdb/ChangeLog:
* ada-valprint.c (val_print_packed_array_elements): Delete
variable "len". Add a type-length check when comparing two
consecutive elements of the array. Use the element's actual
length in call to value_contents_eq.
* ada-lang.c (ada_value_primitive_packed_val): Always return
a value whose type has been resolved.
Consider the following declarations:
subtype Small_Type is Integer range 0 .. 10;
type Record_Type (I : Small_Type := 0) is record
S : String (1 .. I);
end record;
A2 : Array_Type := (1 => (I => 2, S => "AB"),
2 => (I => 1, S => "A"),
3 => (I => 0, S => <>));
Compiled with -fgnat-encodings=minimal, and trying to print
one element of our array, valgrind reports an invalid memory
access. On certain GNU/Linux boxes, malloc even reports it as
well, and causes GDB to crash.
(gdb) print a2(1)
*** glibc detected *** /[...]/gdb:
malloc(): memory corruption: 0x0a30ba48 ***
[crash]
The invalid memory access occurs because of a simple buffer
overflow in ada_value_primitive_packed_val. When this function
is called, it is given a bit_size of 128 (or 16 bytes), which
corresponds to the stride of our array. But the actual size of
each element depends on its value. In particular, A2(1) is a record
whose size is only 6 bytes.
What happens in our example is that we start building a new value
(v) where the element is to be unpacked, with any of its dynamic
properties getting resolved as well. We then unpack the data into
this value's buffer:
unpacked = (unsigned char *) value_contents (v);
[...]
nsrc = len;
[...]
while (nsrc > 0)
{
[...]
unpacked[targ] = accum & ~(~0L << HOST_CHAR_BIT);
[...]
targ += delta;
[...]
nsrc -= 1;
[...]
}
In the loop above, targ starts at zero (for LE architectures),
and len is 16. With delta being +1, we end up iterating 16 times,
writing 16 bytes into a 6-bytes buffer.
This patch fixes the issue by adjusting BIT_SIZE and recomputing
LEN after having resolved our type if the resolved type turns out
to be smaller than bit_size.
gdb/ChangeLog:
* ada-lang.c (ada_value_primitive_packed_val): Recompute
BIT_SIZE and LEN if the size of the resolved type is smaller
than BIT_SIZE * HOST_CHAR_BIT.
Consider the following (Ada) array...
A1 : Array_Type := (1 => (I => 0, S => <>),
2 => (I => 1, S => "A"),
3 => (I => 2, S => "AB"));
... where Array_Type is declared as follow:
subtype Small_Type is Integer range 0 .. 10;
type Record_Type (I : Small_Type := 0) is record
S : String (1 .. I);
end record;
type Array_Type is array (Integer range <>) of Record_Type;
Trying to print the value of each element individually does not
always work. Printing the value of the first one does:
(gdb) p a1(1)
$1 = (i => 0, s => "")
But printing the value of the subsequent ones often does not.
For instance:
(gdb) p a1(2)
$2 = (i => 1, s => "") <<<--- s should be "A"
(gdb) p a1(3)
$3 = (i => 2, s => "") <<<--- s should be "AB"
I traced the problem to ada_value_primitive_packed_val,
which is trying to perform the array subscripting by
extracting the value of the corresponding array element
into a buffer where the contents is now byte-aligned.
The element type that ada_value_primitive_packed_val gets passed
is a dynamic type. As it happens, that dynamic type can get resolved
thanks to:
v = value_at (type, value_address (obj));
type = value_type (v);
However, obj represents the array, so the address given in the call
to value_at represents the value of the first element. As a result,
the solution of component S's upper bound always gets resolved based
on the value of component I in the first element of the array, whose
value is 0, thus leading to GDB mistakely resolving the element type
where S's upper bound is always 0.
The proper fix would be to systematically resolve the element type
first. But, this requires us to extract-and-realign the element's
value so as to be able to pass it as "valaddr" to resolve_dynamic_type.
In the meantime, it's easy to make the situation a little better by
passing "value_address (obj) + offset" as the object address. This
only works when BIT_OFFSET is nul, but that should be the case when
the element type is anything but a scalar, which seems to be the only
situation where it seems important to resolve the type now. And we're
not that worse off otherwise.
But we'll try to find a better solution in a separate patch.
gdb/ChangeLog:
* ada-lang.c (ada_value_primitive_packed_val): Use a more
correct address in call to value_at. Adjust call to
value_address accordingly.
This is another required step towards trying to print the value of
an array of variant records. For instance:
A1 : Array_Type := (1 => (I => 0, S => <>),
2 => (I => 1, S => "A"),
3 => (I => 2, S => "AB"));
... where Array_Type is an array of records whose size is variable:
subtype Small_Type is Integer range 0 .. 10;
type Record_Type (I : Small_Type := 0) is record
S : String (1 .. I);
end record;
type Array_Type is array (Integer range <>) of Record_Type;
What happens is that the ada-valprint modules gets passed an array
whose element type is not resolved yet (since each element of the
array needs to be resolved separately). the module then recurses,
and eventually gets called with the first element of the array.
But because the element hasn't been resolved yet, we end up having
trouble printing its value soon after.
This patch fixes the issue by calling resolve_dynamic_type before
trying to print it.
With this patch, GDB is finally able to print the complete value
for variable "A1":
(gdb) p a1
$1 = ((i => 0, s => ""), (i => 1, s => "A"), (i => 2, s => "AB"))
gdb/ChangeLog:
* ada-valprint.c (ada_val_print_1): Resolve TYPE before trying
to print it.
This is the second part of enhancing the debugger to print the value
of arrays of records whose size is variable when only standard DWARF
info is available (no GNAT encoding). For instance:
subtype Small_Type is Integer range 0 .. 10;
type Record_Type (I : Small_Type := 0) is record
S : String (1 .. I);
end record;
type Array_Type is array (Integer range <>) of Record_Type;
A1 : Array_Type := (1 => (I => 0, S => <>),
2 => (I => 1, S => "A"),
3 => (I => 2, S => "AB"));
Currently, GDB prints the following output:
(gdb) p a1
$1 = (
The error happens while the ada-valprint module is trying to print
the value of an element of our array. Because of the fact that
the array's element (type Record_Type) has a variant size, the DWARF
info for our array provide the array's stride:
<1><749>: Abbrev Number: 10 (DW_TAG_array_type)
<74a> DW_AT_name : (indirect string, offset: 0xb6d): pck__T18s
<74e> DW_AT_byte_stride : 16
<74f> DW_AT_type : <0x6ea>
And because our array has a stride, ada-valprint treats it the same
way as packed arrays (see ada-valprint.c::ada_val_print_array):
if (TYPE_FIELD_BITSIZE (type, 0) > 0)
val_print_packed_array_elements (type, valaddr, offset_aligned,
0, stream, recurse,
original_value, options);
The first thing that we should notice in the call above is that
the "valaddr" buffer and the associated offset (OFFSET_ALIGNED)
is passed, but that the corresponding array's address is not.
This can be explained by looking inside val_print_packed_array_elements,
where we see that the function unpacks each element of our array from
the buffer alone (ada_value_primitive_packed_val), and then prints
the resulting artificial value instead:
v0 = ada_value_primitive_packed_val (NULL, valaddr + offset,
(i0 * bitsize) / HOST_CHAR_BIT,
(i0 * bitsize) % HOST_CHAR_BIT,
bitsize, elttype);
[...]
val_print (elttype, value_contents_for_printing (v0),
value_embedded_offset (v0), 0, stream,
recurse + 1, v0, &opts, current_language);
Of particular interest, here, is the fact that we call val_print
with a null address, which is OK, since we're providing a buffer
instead (value_contents_for_printing). Also, providing an address
might not always possible, since packing could place elements at
boundaries that are not byte-aligned.
Things go south when val_print tries to see if there is a pretty-printer
that could be applied. In particular, one of the first things that
the Python pretty-printer does is to create a value using our buffer,
and the given address, which in this case is null (see call to
value_from_contents_and_address in gdbpy_apply_val_pretty_printer).
value_from_contents_and_address, in turn immediately tries to resolve
the type, using the given address, which is null. But, because our
array element is a record containing an array whose bound is the value
of one of its elements (the "s" component), the debugging info for
the array's upper bound is a reference...
<3><71a>: Abbrev Number: 7 (DW_TAG_subrange_type)
<71b> DW_AT_type : <0x724>
<71f> DW_AT_upper_bound : <0x703>
... to component "i" of our record...
<2><703>: Abbrev Number: 5 (DW_TAG_member)
<704> DW_AT_name : i
<706> DW_AT_decl_file : 2
<707> DW_AT_decl_line : 6
<708> DW_AT_type : <0x6d1>
<70c> DW_AT_data_member_location: 0
... where that component is located at offset 0 of the start
of the record. dwarf2_evaluate_property correctly determines
the offset where to load the value of the bound from, but then
tries to read that value from inferior memory using the address
that was given, which is null. See case PROP_ADDR_OFFSET in
dwarf2_evaluate_property:
val = value_at (baton->offset_info.type,
pinfo->addr + baton->offset_info.offset);
This triggers a memory error, which then causes the printing to terminate.
Since there are going to be situations where providing an address
alone is not going to be sufficient (packed arrays where array elements
are not stored at byte boundaries), this patch fixes the issue by
enhancing the type resolution to take both address and data. This
follows the same principle as the val_print module, where both
address and buffer ("valaddr") can be passed as arguments. If the data
has already been fetched from inferior memory (or provided by the
debugging info in some form -- Eg a constant), then use that data
instead of reading it from inferior memory.
Note that this should also be a good step towards being able to handle
dynamic types whose value is stored outside of inferior memory
(Eg: in a register).
With this patch, GDB isn't able to print all of A1, but does perform
a little better:
(gdb) p a1
$1 = ((i => 0, s => , (i => 1, s => , (i => 2, s => )
There is another issue which is independent of this one, and will
therefore be patched separately.
gdb/ChangeLog:
* dwarf2loc.h (struct property_addr_info): Add "valaddr" field.
* dwarf2loc.c (dwarf2_evaluate_property): Add handling of
pinfo->valaddr.
* gdbtypes.h (resolve_dynamic_type): Add "valaddr" parameter.
* gdbtypes.c (resolve_dynamic_struct): Set pinfo.valaddr.
(resolve_dynamic_type_internal): Set pinfo.valaddr.
Add handling of addr_stack->valaddr.
(resolve_dynamic_type): Add "valaddr" parameter.
Set pinfo.valaddr field.
* ada-lang.c (ada_discrete_type_high_bound): Update call to
resolve_dynamic_type.
(ada_discrete_type_low_bound): Likewise.
* findvar.c (default_read_var_value): Likewise.
* value.c (value_from_contents_and_address): Likewise.
Consider the following (Ada) variable...
A1 : Array_Type := (1 => (I => 0, S => <>),
2 => (I => 1, S => "A"),
3 => (I => 2, S => "AB"));
... where Array_Type is an array of records whose size is variable:
subtype Small_Type is Integer range 0 .. 10;
type Record_Type (I : Small_Type := 0) is record
S : String (1 .. I);
end record;
type Array_Type is array (Integer range <>) of Record_Type;
Trying to print the value of this array currently results in the following
error:
(gdb) p a1
Cannot access memory at address 0x61c000
What happens in this case, is that the compiler describes our array
as an array with a specific stride (and bounds being static 1..3):
<1><749>: Abbrev Number: 10 (DW_TAG_array_type)
<74a> DW_AT_name : (indirect string, offset: 0xb6d): pck__T18s
<74e> DW_AT_byte_stride : 16
<74f> DW_AT_type : <0x6ea>
<2><757>: Abbrev Number: 11 (DW_TAG_subrange_type)
<758> DW_AT_type : <0x75e>
<75c> DW_AT_upper_bound : 3
This is because we cannot use, in this case, the size of the record
to determine that stride, since the size of the record depends on
its contents. So the compiler helps us by providing that stride.
The problems start when trying to resolve that type. Because the elements
contained in that array type are dynamic, the array itself is considered
dynamic, and thus we end up creating a resolved version of that array.
And during that resolution, we were not handling the case where the array
had a stride. See gdbtypes.c::resolve_dynamic_array...
return create_array_type (copy_type (type),
elt_type,
range_type);
As a result, we created an array whose stride was based on the size
of elt_type, which a record whose size isn't static and irrelevant
regardless.
This patch fixes is by calling create_array_type_with_stride instead.
As it happens, there is another issue for us to be able to print
the value of our array, but those are independent of this patch
and will be handled separately. For now, the patch allows us to
get rid of the first error, and the output is now:
(gdb) p a1
$1 = (
gdb/ChangeLog:
* gdbtypes.c (resolve_dynamic_array): Use
create_array_type_with_stride instead of create_array_type.
gas * config/tc-msp430.c (MAX_OP_LEN): Increase to 4096.
(msp430_make_init_symbols): New function.
(msp430_section): Call it.
(msp430_frob_section): Likewise.
ld * emulparams/msp430elf.sh (TEMPLATE_NAME): Change to msp430.
* scripttempl/msp430.sc (.text): Add .lower.text and .either.text.
(.data): Add .lower.data and .either.data.
(.bss): Add .lower.bss and .either.bss.
(.rodata): Add .lower.rodata and .either.rodata.
* emultempl/msp430.em: New file. Implements a new orphan
placement algorithm that divides sections between lower and upper
memory regions.
* Makefile.am (emsp430elf.c): Depend upon msp430.em.
*emsp430X.c): Likewise.
* Makefine.in: Regenerate.
Currently every fixup in the current segment is checked when relaxing
trampoline frag. This is very expensive. Make a searchable array of
fixups pointing at potentially oversized jumps at the beginning of every
relaxation pass and only check subset of this cache in the reach of
single jump from the trampoline frag currently being relaxed.
Original profile:
% time self children called name
-----------------------------------------
370.16 593.38 12283048/12283048 relax_segment
98.4 370.16 593.38 12283048 xtensa_relax_frag
58.91 269.26 2691463834/2699602236 xtensa_insnbuf_from_chars
68.35 68.17 811266668/813338977 S_GET_VALUE
36.85 29.51 2684369246/2685538060 xtensa_opcode_decode
28.34 8.84 2684369246/2685538060 xtensa_format_get_slot
12.39 5.94 2691463834/2699775044 xtensa_format_decode
0.03 4.60 4101109/4101109 relax_frag_for_align
0.18 1.76 994617/994617 relax_frag_immed
0.07 0.09 24556277/24851220 new_logical_line
0.06 0.00 12283048/14067410 as_where
0.04 0.00 7094588/15460506 xtensa_format_num_slots
0.00 0.00 1/712477 xtensa_insnbuf_alloc
-----------------------------------------
Same data, after optimization:
% time self children called name
-----------------------------------------
0.51 7.47 12283048/12283048 relax_segment
58.0 0.51 7.47 12283048 xtensa_relax_frag
0.02 4.08 4101109/4101109 relax_frag_for_align
0.18 1.39 994617/994617 relax_frag_immed
0.01 0.98 555/555 xtensa_cache_relaxable_fixups
0.21 0.25 7094588/16693271 xtensa_insnbuf_from_chars
0.06 0.12 24556277/24851220 new_logical_line
0.06 0.00 7094588/15460506 xtensa_format_num_slots
0.02 0.04 7094588/16866079 xtensa_format_decode
0.05 0.00 12283048/14067410 as_where
0.00 0.00 1/712477 xtensa_insnbuf_alloc
0.00 0.00 93808/93808 xtensa_find_first_cached_fixup
-----------------------------------------
2015-05-02 Max Filippov <jcmvbkbc@gmail.com>
gas/
* config/tc-xtensa.c (cached_fixupS, fixup_cacheS): New typedefs.
(struct cached_fixup, struct fixup_cache): New structures.
(fixup_order, xtensa_make_cached_fixup),
(xtensa_realloc_fixup_cache, xtensa_cache_relaxable_fixups),
(xtensa_find_first_cached_fixup, xtensa_delete_cached_fixup),
(xtensa_add_cached_fixup): New functions.
(xtensa_relax_frag): Cache fixups pointing at potentially
oversized jumps at the beginning of every relaxation pass. Only
check subset of this cache in the reach of single jump from the
trampoline frag currently being relaxed.
