Fix a NULL dereference seen when assembling invalid input.
PR 27198
* config/tc-i386.c (need_plt32_p): Return FALSE for NULL symbol.
* testsuite/gas/i386/pr27198.d,
* gas/testsuite/gas/i386/pr27198.err,
* gas/testsuite/gas/i386/pr27198.s: New test.
* gas/testsuite/gas/i386/i386.exp: Run it.
When deciding if a single member comdat group section in file FOO should
be discarded by a linkonce section in file BAR, we check if 2 sections
define the same set of local and global symbols. When only one of the
files doesn't contain the unused section symbols in the symbol table,
such as object files generated by clang or GNU assembler with
commit d1bcae833b
Author: H.J. Lu <hjl.tools@gmail.com>
Date: Thu Jan 7 06:42:00 2021 -0800
ELF: Don't generate unused section symbols
the check will fail since one file has the extra unused section symbols.
We should ignore both undefined and section symbols in the symbol table
when making such a decision.
bfd/
PR ld/27193
* elflink.c (elf_create_symbuf): Also ignore section symbols.
ld/
PR ld/27193
* testsuite/ld-i386/i386.exp: Run PR ld/27193 test.
* testsuite/ld-i386/pr27193.dd: New file.
* testsuite/ld-i386/pr27193a.o.bz2: Likewise.
* testsuite/ld-i386/pr27193b.s: Likewise.
On Fedora 33 x86-64 with glibc 2.32-3, ifuncmain6pie failed with:
./ifuncmain6pie: IFUNC symbol 'foo' referenced in './ifuncmod6.so' is defined in the executable and creates an unsatisfiable circular dependency.
FAIL ifuncmain6pie (exit status: 127)
Remove non-JUMP_SLOT relocations against foo in ifuncmod6.so, which
trigger the circular IFUNC dependency.
* testsuite/ifuncmain6pie.c: Remove non-JUMP_SLOT relocations
against foo in ifuncmod6.so.
* testsuite/ifuncmod6.c: Likewise.
On x86, glibc 2.33 starts to issue a fatal error message when calling
IFUNC function defined in the unrelocated executable from a shared
library.
1. Update x86 ELF linker to always convert IFUNC function defined in
position-dependent executable (PDE) to the normal function. GOT in PDE
will be updated by R_*_IRELATIVE at run-time.
2. Update PR ld/23169 tests not to compare function address of external
IFUNC function in the shared object to avoid calling the IFUNC function
defined in the unrelocated executable.
3. Remove pr23169e tests which call the IFUNC function defined in the
unrelocated position-independent executable from a shared library.
bfd/
PR ld/23169
* elfxx-x86.c (_bfd_x86_elf_link_fixup_ifunc_symbol): Don't
check pointer_equality_needed.
ld/
PR ld/23169
* testsuite/ld-ifunc/ifunc.exp: Replace pr23169c.rd with
pr23169a.rd for pr23169c and pr23169f. Remove pr23169e tests.
* testsuite/ld-ifunc/pr23169a.c (foo): Don't compare function
address.
So that no one need worry about the value of Z_OK.
bfd/
* compress.c (decompress_contents): Tidy inflateEnd result test.
binutils/
* readelf.c (uncompress_section_contents): Tidy inflateEnd result test.
This patch makes undefined unversioned dynamic symbols use
VER_NDX_GLOBAL (version 1) rather than VER_NDX_LOCAL (version 0).
There really isn't much use for an undefined local dynamic symbol, so
we may as well use the logically correct value in .gnu.version.
https://docs.oracle.com/cd/E26505_01/html/E26506/chapter6-54676.html
bfd/
PR 26002
* elflink.c (elf_link_output_extsym): Use version 1 in
.gnu.version for undefined unversioned symbols.
ld/
PR 26002
* testsuite/ld-elfvers/vers6.dsym: Expect "Base" for undefined
unversioned symbols.
* testsuite/ld-elfvers/vers16.dsym: Likewise.
Now that all port tests live under testsuite/sim/*/, and none live
in testsuite/ directly, flatten the structure by moving all of the
dirs under testsuite/sim/ to testsuite/ directly.
We need to stop passing --tool to dejagnu so that it searches all
dirs and not just ones that start with "sim". Since we have no
other dirs in this tree, and no plans to add any, should be fine.
Compare pointers to nullptr, not 0. I also fixed a trailing
whitespace in the same function.
There should be no user visible changes after this commit.
gdb/ChangeLog:
* tui/tui.c (tui_is_window_visible): Compare to nullptr, not 0.
When running test-case gdb.fortran/array-slices.exp with target board
unix/-m32, we run into:
...
(gdb) print /x &array4d^M
$69 = 0xffffb620^M
(gdb) print /x (&array4d) + sizeof (array4d)^M
$70 = 0x95c620^M
(gdb) FAIL: gdb.fortran/array-slices.exp: repack=on: test 9: check sizes match
...
The expressions calculate the start and end of an array, but the calculation
of the end expression has an unexpected result (given that it lies before the
start of the array). By printing "sizeof (array4d)" as a separate
expression:
...
(gdb) print /x sizeof (array4d)
$1 = 0xc40
...
it becomes clear we expected to get 0xffffb620 + 0xc40 == 0xffffc260 instead.
The problem is that using the '&' returns a pointer type:
...
(gdb) p &array4d
$5 = (PTR TO -> ( integer(kind=4) (-3:3,7:10,-3:3,-10:-7) )) 0xffffbe00
...
which has the consequence that the addition is done as pointer arithmetic.
Fix this by using the result of "print /x &array4d" instead of &array4d in the
addition.
Tested on x86_64-linux.
gdb/testsuite/ChangeLog:
2021-01-15 Tom de Vries <tdevries@suse.de>
PR testsuite/26997
* gdb.fortran/array-slices.exp (run_test): Avoid pointer arithmetic
when adding sizeof.
bfd/
* elfnn-riscv.c: Indent, labels and GNU coding standards tidy,
also aligned the code.
gas/
* config/tc-riscv.c: Indent and GNU coding standards tidy,
also aligned the code.
* config/tc-riscv.h: Likewise.
include/
* opcode/riscv.h: Indent and GNU coding standards tidy,
also aligned the code.
opcodes/
* riscv-opc.c (riscv_gpr_names_abi): Aligned the code.
(riscv_fpr_names_abi): Likewise.
(riscv_opcodes): Likewise.
(riscv_insn_types): Likewise.
Error and warning messages usually starting with lower case letter,
and without the period at the end. Besides, add the prefixed "internel:"
at the beginning of the messages when they are caused internally.
Also fix indents and typos.
bfd/
* elfnn-riscv.c (riscv_merge_attributes): Fix typos of messages.
gas/
* config/tc-riscv.c: Error and warning messages tidy.
* testsuite/gas/riscv/priv-reg-fail-fext.l: Updated.
* testsuite/gas/riscv/priv-reg-fail-read-only-01.l: Likewise.
* testsuite/gas/riscv/priv-reg-fail-read-only-02.l: Likewise.
* testsuite/gas/riscv/priv-reg-fail-rv32-only.l: Likewise.
* testsuite/gas/riscv/priv-reg-fail-version-1p10.l: Likewise.
* testsuite/gas/riscv/priv-reg-fail-version-1p11.l: Likewise.
* testsuite/gas/riscv/priv-reg-fail-version-1p9p1.l: Likewise.
ld/
* testsuite/ld-riscv-elf/attr-merge-priv-spec-failed-01.d: Updated.
* testsuite/ld-riscv-elf/attr-merge-priv-spec-failed-02.d: Likewise.
* testsuite/ld-riscv-elf/attr-merge-priv-spec-failed-03.d: Likewise.
* testsuite/ld-riscv-elf/attr-merge-priv-spec-failed-04.d: Likewise.
* testsuite/ld-riscv-elf/attr-merge-priv-spec-failed-05.d: Likewise.
* testsuite/ld-riscv-elf/attr-merge-priv-spec-failed-06.d: Likewise.
opcodes/
* riscv-dis.c (parse_riscv_dis_option): Fix typos of message.
The GNU coding standards said the comments should be complete sentences
and end with a period and two spaces. But sometimes it should be more
cleaner when the comments only include a word or codes. Therefore, I made
the following changes after referring to other target/generic codes,
* Try to write sentences in comments, must end with a period and two spaces.
* End with two spaces without a period for codes/instructions only.
* End with one space without a period for a single word/variable only.
Besids, also rewrite/remove some comments which are obsolete or too long,
and fix indents for comments.
bfd/
* elfnn-riscv.c: Comments tidy and improvement.
* elfxx-riscv.c: Likewise.
* elfxx-riscv.h: Likewise.
gas/
* config/tc-riscv.c: Comments tidy and improvement. Also update
comment "fallthru" to "Fall through" that end with a period and
two spaces.
include/
* elf/riscv.h: Comments tidy and improvement.
* opcode/riscv-opc.h: Likewise.
* opcode/riscv.h: Likewise.
opcodes/
* riscv-dis.c: Comments tidy and improvement.
* riscv-opc.c: Likewise.
Now that we've moved all ports to dejagnu & testsuite/sim/, the only
thing the testsuite/configure script has been doing is filling in the
sim_arch field in the testsuite/Makefile. We can simply let the top
sim/configure script do that for us now. This simplifies & speeds up
the build a bit by killing an entire configure script.
This is the only target using a dir directly under testsuite/. All
others use sim/<arch>/ instead. Relocate it so all targets look the
same, and so we can leverage the common test harness.
We drop loop.s in the process because it was never referenced and
was just 2 lines of code.
All other test files are moved & have directives added to the top so
that the test harness can invoke them correctly.
No tests were ever added in here in the ~22 years since it was first
created. Seems unlikely any tests will be added at this rate, and
the sim/mips/ testdir already has some (light) coverage for this
target. So punt the tree.
The frv-elf subdir contained five tests:
* cache: A cache test of some sort.
* exit47: A program to test exit status of 47 from sim.
* grloop: Some basic limited loop test program.
* hello: Standard "hello world" output program.
* loop: An infinite loop program.
The loop.s test is never referenced anywhere, and is all of 2 lines.
Anyone who really needs a while(1); test case and re-implement it
themselves locally.
The cache.s code isn't referenced anywhere because it requires some
custom args to the run program, and when this testcase was added, we
didn't have any support for that. We do now, so we can add a header
to enable that. Turns out the code crashes even with those, so turn
around and mark it xfail. Maybe someone someday will care.
That leaves the small exit47, grloop, and hello tests. Now that the
sim test harness supports testing for custom exit status, we can move
them all to sim/frv/ to maintain test coverage.
The remaining differences between frv-elf & sim/frv are:
* frv-elf/ runs for frv-*-elf while sim/frv/ runs for frv*-*-*.
* frv-elf/ runs "*.s" files while sim/frv/ only has .cgs and such.
On closer inspection, these are also meaningless distinctions:
* There is nothing specific to the tests that require an *-elf
target. Normally that would mean newlib+libgloss type stuff,
but there's no such requirement in frv-elf/.
* The ".s" suffix is the standard "this is an assembly file" suffix.
Since FRV is a CGEN target, we can reuse the existing convention of
".ms" to mean "miscellaneous .s" as in "this is an assembly file,
and run/bucket its test results in the miscellaneous category".
So moving frv-elf/{cache,exit47,grloop,hello}.s to sim/frv/*.ms makes
sense and simplifies things quite a bit for the target while also
slightly increasing the coverage for some tuples.
The m32r-elf subdir contained three tests:
* exit47: A program to test exit status of 47 from sim.
* hello: Standard "hello world" output program.
* loop: An infinite loop program.
There's already a sim/m32r/hello.ms test that does exactly the same
thing as m32r-elf/hello.s, so we can delete that.
The loop.s test is never referenced anywhere, and is all of 2 lines.
Anyone who really needs a while(1); test case and re-implement it
themselves locally.
That leaves the single exit47 test. Now that the sim test harness
supports testing for custom exit status, we can easily move that to
sim/m32r/exit47.ms to maintain test coverage.
The remaining differences between m32r-elf & sim/m32r are:
* m32r-elf/ runs for m32r-*-elf while sim/m32r/ runs for m32r*-*-*.
* m32r-elf/ runs "*.s" files while sim/m32r/ runs "*.ms" files.
On closer inspection, these are also meaningless distinctions:
* There is nothing specific to the tests that require an *-elf
target. Normally that would mean newlib+libgloss type stuff,
but there's no such requirement in m32r-elf/.
* The ".s" suffix is the standard "this is an assembly file"
suffix. Turns out ".ms" is just how sim/m32r/ (and a few other
CGEN based targets) categorize/bucket test cases. It simply
means "miscellaneous .s" as in "this is an assembly file, and
run/bucket its test results in the miscellaneous category".
So moving m32r-elf/exit47.s to sim/m32r/exit47.ms makes sense and
simplifies things quite a bit for the target while also slightly
increasing the coverage for some tuples.
Some tests want to verify they can control the exit status, and
allowing any non-zero value would allow tests to silently fail:
if it crashed & exited 1, or forced all non-zero to 1, then we
wouldn't be able to differentiate with a test exiting with a
status like 47.
Extend the test harness to allow tests to declare their expected
exit status that would be defined as a "pass". This requires a
small tweak to the sim_run API to return the status directly, but
that shouldn't be a big deal as it's only used by sim code.
Since is_trivially_default_constructible was not implemented before gcc-5
it cannot be used with gcc-4.x.
Fix the build by using conditional compilation around that line.
Use the equivalent is_trivially_constructible<T> instead, since
we already have HAVE_IS_TRIVIALLY_CONSTRUCTIBLE for that purpose.
Fixes: 098caef485 ("Refactor struct trad_frame_saved_regs")
gdb:
2021-01-14 Bernd Edlinger <bernd.edlinger@hotmail.de>
* trad-frame.c (trad_frame_alloc_saved_regs): Avoid compile-error
because is_trivially_default_constructible was first implemented with
gcc-5.
Linker should never generate dynamic relocations for relocations in
non-SEC_ALLOC section which has no impact on run-time behavior. Such
relocations should be resolved to 0.
PR ld/26688
* elf32-bfin.c (bfinfdpic_relocate_section): Skip non SEC_ALLOC
section for R_BFIN_FUNCDESC.
Consider the following test-case small.c:
...
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main (void) {
int *p = (int *)malloc (sizeof(int) * 4);
memset (p, 0, sizeof(p));
printf ("p[0] = %d; p[3] = %d\n", p[0], p[3]);
return 0;
}
...
On Ubuntu 20.04, we get:
...
$ gcc -O0 -g small.c
$ gdb -batch a.out -ex start -ex step
Temporary breakpoint 1, main () at small.c:6
6 int *p = (int *) malloc(sizeof(int) * 4);
p[0] = 0; p[3] = 0
[Inferior 1 (process $dec) exited normally]
...
but after switching off the on-by-default fcf-protection, we get the desired
behaviour:
...
$ gcc -O0 -g small.c -fcf-protection=none
$ gdb -batch a.out -ex start -ex step
Temporary breakpoint 1, main () at small.c:6
6 int *p = (int *) malloc(sizeof(int) * 4);
7 memset (p, 0, sizeof(p));
...
Using "set debug infrun 1", the first observable difference between the two
debug sessions is that with -fcf-protection=none we get:
...
[infrun] process_event_stop_test: stepped into dynsym resolve code
...
In this case, "in_solib_dynsym_resolve_code (malloc@plt)" returns true because
"in_plt_section (malloc@plt)" returns true.
With -fcf-protection=full, "in_solib_dynsym_resolve_code (malloc@plt)" returns
false because "in_plt_section (malloc@plt)" returns false, because the section
name for malloc@plt is .plt.sec instead of .plt, which is not handled in
in_plt_section:
...
static inline int
in_plt_section (CORE_ADDR pc)
{
return pc_in_section (pc, ".plt");
}
...
Fix this by handling .plt.sec in in_plt_section.
Tested on x86_64-linux.
[ Another requirement to be able to reproduce this is to have a dynamic linker
with a "malloc" minimal symbol, which causes find_solib_trampoline_target to
find it, such that skip_language_trampoline returns the address for the
dynamic linkers malloc. This causes the step machinery to set a breakpoint
there, and to continue, expecting to hit it. Obviously, we execute glibc's
malloc instead, so the breakpoint is not hit and we continue to program
completion. ]
gdb/ChangeLog:
2021-01-14 Tom de Vries <tdevries@suse.de>
PR breakpoints/27151
* objfiles.h (in_plt_section): Handle .plt.sec.
When running test-case gdb.base/style.exp with target board unix/-m32, we run
into (stripped styling from output, shortened file name):
...
(gdb) frame
argv=0xffffc714) at src/gdb/testsuite/gdb.base/style.c:45
45 return some_called_function (); /* break here */
(gdb) FAIL: gdb.base/style.exp: frame when width=20
...
while with native we have instead:
...
(gdb) frame
argv=0x7fffffffd478)
at src/gdb/testsuite/gdb.base/style.c:45
45 return some_called_function (); /* break here */
(gdb) PASS: gdb.base/style.exp: frame when width=20
...
The problem is that due to argv having a different length for -m32, we get a
different layout, and the test-case doesn't accommodate for that.
Fix this by using a different regexp depending on the length of argv.
Tested on x86_64-linux.
gdb/testsuite/ChangeLog:
2021-01-14 Tom de Vries <tdevries@suse.de>
PR testsuite/24590
* gdb.base/style.exp: Handle shorter argv in frame command output.
The Blackfin ELF compiler requires the user to explicitly select a CPU
target else it will fail:
bfin-elf-gcc: error: no processor type specified for linking
Select the sim target for these tests since we should (hopefully) have
access to the simulator. At least, it's more likely than having access
to a real development board.
This makes the pass/fail numbers increase by a lot:
-# of expected passes 398
-# of unexpected failures 6
+# of expected passes 587
+# of unexpected failures 109
It looks like the vast majority of new failures are due to our omission
of COPY relocations:
/* Bfin does not currently have a COPY reloc. */
if ((h->root.u.def.section->flags & SEC_ALLOC) != 0)
{
_bfd_error_handler (_("the bfin target does not currently support the generation of copy relocations"));
return FALSE;
}
There doesn't seem to be a way to easily disable tests that cause copy
relocations though, lets just take the hit for now.
* testsuite/config/default.exp [bfin*-elf*] (gcc_B_opt): Append -msim.
The makefile has comments about old versions of bison/yacc generating
warnings, but that doesn't apply to the lexer which comes from flex.
As far as I can tell, the warnings in the Blackfin lexer can be fixed
with defines that have been supported back through flex in 2002. So
lets turn on -Werror for it and see if anyone notices. If they do,
they can report their exact tool versions so we can record that here.
The h8300 sim has its own implementation for memory handling that I'd
like to replace with the common sim memory code. However, it's got a
weird bit of code it calls "eightbit mem" that makes this not as easy
as it would otherwise be. The code has this comment:
/* These define the size of main memory for the simulator.
Note the size of main memory for the H8/300H is only 256k. Keeping it
small makes the simulator run much faster and consume less memory.
The linker knows about the limited size of the simulator's main memory
on the H8/300H (via the h8300h.sc linker script). So if you change
H8300H_MSIZE, be sure to fix the linker script too.
Also note that there's a separate "eightbit" area aside from main
memory. For simplicity, the simulator assumes any data memory reference
outside of main memory refers to the eightbit area (in theory, this
can only happen when simulating H8/300H programs). We make no attempt
to catch overlapping addresses, wrapped addresses, etc etc. */
I've read the H8/300 Programming Manual and the H8/300H Software Manual
and can't find documentation on it. The closest I can find is the bits
about the exception vectors, but that sounds like a convention where the
first 256 bytes of memory are used for a special purpose. The sim will
actually allocate a sep memory buffer of 256 bytes and you address it by
accessing anywhere outside of main memory. e.g. I would expect code to
access it like:
uint32_t *data = (void *)0;
data[0] = reset_exception_vector;
not like the sim expects like:
uint8_t *data = (void *)0x1000000;
data[0] = ...;
The gcc manual has an "eightbit_data" attribute:
Use this attribute on the H8/300, H8/300H, and H8S to indicate that
the specified variable should be placed into the eight-bit data
section. The compiler generates more efficient code for certain
operations on data in the eight-bit data area. Note the eight-bit
data area is limited to 256 bytes of data.
And the gcc code implies that it's accessed via special addressing:
eightbit_data: This variable lives in the 8-bit data area and can
be referenced with 8-bit absolute memory addresses.
I'm fairly certain these are referring to the 8-bit addressing modes
that allow access to 0xff00 - 0xffff with only an 8-bit immediate.
They aren't completely separate address spaces which this eightbit
memory buffer occupies.
But the sim doesn't access its eightbit memory based on specific insns,
it does it purely on the addresses requested.
Unfortunately, much of this code was authored by Michael Snyder, so I
can't ask him :(. I asked Renesas support and they didn't know:
https://renesasrulz.com/the_vault/f/archive-forum/6952/question-about-eightbit-memory
So I've come to the conclusion that this was a little sim-specific hack
done for <some convenience> and has no relation to real hardware. And
as such, let's drop it until someone notices and can provide a reason
for why we need to support it.
The default watchpoint handler is NULL. That means any port that
sets the STATE_WATCHPOINTS->pc field will crash if you try to use
the --watch options but don't configure the interrupt handler. In
the past, you had to setup STATE_WATCHPOINTS->pc if you wanted to
support PC profiling, and while that was fixed a while ago, we have
a lot of ports who still configure it.
We already add a default set of interrupts (just "int") if the port
doesn't define any. Let's also add a default handler that raises a
SIGTRAP. When connected to gdb, this is a breakpoint which is what
people would expect. When running standalone, it'll abort the sim,
but it's unclear whether there's anything better to do there. This
really is just to make the watchpoint module more usable out of the
box for most ports with very little setup, at least inside of gdb.
This commit builds on work started in the following two commits:
commit 24ed6739b6
Date: Thu Jan 30 14:35:40 2020 +0000
gdb/remote: Restore support for 'S' stop reply packet
commit cada5fc921
Date: Wed Mar 11 12:30:13 2020 +0000
gdb: Handle W and X remote packets without giving a warning
This is related to how GDB handles remote targets that send back 'S'
packets.
In the first of the above commits we fixed GDB's ability to handle a
single process, single threaded target that sends back 'S' packets.
Although the 'T' packet would always be preferred to 'S' these days,
there's nothing really wrong with 'S' for this situation.
The second commit above fixed an oversight in the first commit, a
single-process, multi-threaded target can send back a process wide
event, for example the process exited event 'W' without including a
process-id, this also is fine as there is no ambiguity in this case.
In PR gdb/26819 we run into yet another problem with the above
commits. In this case we have a single process with two threads, GDB
hits a breakpoint in thread 2 and then performs a stepi:
(gdb) b main
Breakpoint 1 at 0x1212340830: file infinite_loop.S, line 10.
(gdb) c
Continuing.
Thread 2 hit Breakpoint 1, main () at infinite_loop.S:10
10 in infinite_loop.S
(gdb) set debug remote 1
(gdb) stepi
Sending packet: $vCont;s:2#24...Packet received: S05
../binutils-gdb/gdb/infrun.c:5807: internal-error: int finish_step_over(execution_control_state*): Assertion `ecs->event_thread->control.trap_expected' failed.
What happens in this case is that on the RISC-V target displaced
stepping is not supported, so when the stepi is issued GDB steps just
thread 2. As only a single thread was set running the target decides
that is can get away with sending back an 'S' packet without a
thread-id. GDB then associates the stop with thread 1 (the first
non-exited thread), but as thread 1 was not previously set executing
the assertion seen above triggers.
As an aside I am surprised that the target sends pack 'S' in this
situation. The target is happy to send back 'T' (including thread-id)
when multiple threads are set running, so (to me) it would seem easier
to just always use the 'T' packet when multiple threads are in use.
However, the target only uses 'T' when multiple threads are actually
executing, otherwise an 'S' packet it used.
Still, when looking at the above situation we can see that GDB should
be able to understand which thread the 'S' reply is referring too.
The problem is that is that in commit 24ed6739b6 (above) when a stop
reply comes in with no thread-id we look for the first non-exited
thread and select that as the thread the stop applies too.
What we should really do is select the first non-exited, resumed thread,
and associate the stop event with this thread. In the above example
both thread 1 and 2 are non-exited, but only thread 2 is resumed, so
this is what we should use.
There's a test for this issue included which works with stock
gdbserver by disabling use of the 'T' packet, and enabling
'scheduler-locking' within GDB so only one thread is set running.
gdb/ChangeLog:
PR gdb/26819
* remote.c
(remote_target::select_thread_for_ambiguous_stop_reply): New
member function.
(remote_target::process_stop_reply): Call
select_thread_for_ambiguous_stop_reply.
gdb/testsuite/ChangeLog:
PR gdb/26819
* gdb.server/stop-reply-no-thread-multi.c: New file.
* gdb.server/stop-reply-no-thread-multi.exp: New file.
Change-Id: I9b49d76c2a99063dcc76203fa0f5270a72825d15
The previous patch made the commit_resume implementations in the record
targets unnecessary, as the remote target's commit_resume implementation
won't commit-resume threads for which it didn't see a resume. This
patch removes them.
gdb/ChangeLog:
* record-btrace.c (class record_btrace_target): Remove.
(record_btrace_target::commit_resume): Remove.
* record-full.c (class record_full_target): Remove.
(record_full_target::commit_resume): Remove.
Change-Id: I3a68d3d726fb09d8b7165b4edefc330d27803b27
The next patch moves the target commit_resume method to be a
process_stratum_target-only method. The only non-process targets that
currently implement the commit_resume method are the btrace and full
record targets. The only reason they need to do so is to prevent a
commit resume from reaching the beneath (process) target if they are
currently replaying.
This is important if a record target is used on top of the remote target
(the only process target implementing the commit_resume method).
Currently, the remote target checks the `thread_info::executing` flag of
a thread to know if it should commit resume that thread:
if (!tp->executing || remote_thr->vcont_resumed)
continue;
The `tp->executing` flag is set by infrun when it has asked the target
stack to resume the thread, and therefore if the thread is executing,
from its point of view. It _not_ equivalent to whether the remote
target was asked to resume this thread.
Indeed, if infrun asks the target stack to resume some thread while the
record target is replaying, the record target won't forward the resume
request the remote target beneath, because we don't actually want to
resume the thread on the execution target. But the `tp->executing` flag
is still set, because from the point of view of infrun, the thread
executes. So, if the commit_resume call wasn't intercepted by the
record target as it is today and did reach the remote target, the remote
target would say "Oh, this thread should be executing and I haven't
vCont-resumed it! I must vCont-resume it!". But that would be wrong,
because it was never asked to resume this thread, the resume request did
not reach it. This is why the record targets currently need to
implement commit_resume: to prevent the beneath target from
commit_resuming threads it wasn't asked to resume.
Since commit_resume will become a method on process_stratum_target in
the following patch, record targets won't have a chance to intercept the
calls and that would result in the remote target commit_resuming threads
it shouldn't. To avoid this, this patch makes the remote target track
its own thread resumption state. That means, tracking which threads it
was asked to resume via target_ops::resume. Regardless of the context
of this patch, I think this change makes it easier to understand how
resume / commit_resume works in the remote target. It makes the target
more self-contained, as it only depends on what it gets asked to do via
the target methods, and not on tp->executing, which is a flag maintained
from the point of view of infrun.
I initially made it so this state was only used when the remote target
operates in non-stop mode, since commit_resume is only used when the
target is non-stop. However, it's more consistent and it can be useful
to maintain this state even in all-stop too. In all-stop, receiving a
stop notification for one thread means all threads of the target are
considered stopped.
From the point of view of the remote target, there are three states a
thread can be in:
1. not resumed
2. resumed but pending vCont-resume
3. resumed
State 2 only exists when the target is non-stop.
As of this patch, valid state transitions are:
- 1 -> 2 (through the target resume method if in non-stop)
- 2 -> 3 (through the target commit_resume method if in non-stop)
- 1 -> 3 (through the target resume method if in all-stop)
- 3 -> 1 (through a remote stop notification / reporting an event to the
event loop)
A subsequent patch will make it possible to go from 2 to 1, in case
infrun asks to stop a thread that was resumed but not commit-resumed
yet. I don't think it can happen as of now.
In terms of code, this patch replaces the vcont_resumed field with an
enumeration that explicitly represents the three states described above.
The last_resume_sig and last_resume_step fields are moved to a structure
which is clearly identified as only used when the thread is in the
"resumed but pending vCont-resume" state.
gdb/ChangeLog:
* remote.c (enum class resume_state): New.
(struct resumed_pending_vcont_info): New.
(struct remote_thread_info) <resume_state, set_not_resumed,
set_resumed_pending_vcont, resumed_pending_vcont_info,
set_resumed, m_resume_state, m_resumed_pending_vcont_info>:
New.
<last_resume_step, last_resume_sig, vcont_resumed>: Remove.
(remote_target::remote_add_thread): Adjust.
(remote_target::process_initial_stop_replies): Adjust.
(remote_target::resume): Adjust.
(remote_target::commit_resume): Rely on state in
remote_thread_info and not on tp->executing.
(remote_target::process_stop_reply): Adjust.
Change-Id: I10480919ccb4552faa62575e447a36dbe7c2d523
Add the standard arc_debug_printf, but also arc_linux_debug_printf,
arc_linux_nat_debug_printf and arc_newlib_debug_printf to match the
prefixes currently used in the debug messages.
gdb/ChangeLog:
* arc-tdep.h (arc_debug_printf): New.
* arc-tdep.c: Use arc_debug_printf.
* arc-linux-nat.c (arc_linux_nat_debug_printf): Add and use.
* arc-linux-tdep.c (arc_linux_debug_printf): Add and use.
* arc-newlib-tdep.c (arc_newlib_debug_printf): Add and use.
Change-Id: I5d937566ed7a1925f7982e8809802c8f0560d8c6
Shahab suggested we get rid of the verbosity level for the ARC debug
logging [1]. This patch does that, before doing any other change.
gdb/ChangeLog:
* arc-tdep.h (arc_debug): Change type to bool.
* arc-tdep.c (arc_debug): Change type to bool.
(arc_analyze_prologue): Adjust.
(_initialize_arc_tdep): Use add_setshow_boolean_cmd.
* arc-linux-nat.c (ps_get_thread_area): Adjust.
[1] https://sourceware.org/pipermail/gdb-patches/2021-January/175075.html
Change-Id: I16688bd42ed8978ae1acf57012c8d41a943044a5
Bool-ify the return type of maybe_add_script_text and
maybe_add_script_file, the loaded parameter and related things.
gdb/ChangeLog:
* auto-load.c (struct loaded_script) <loaded>: Change to bool.
(maybe_add_script_file): Change return type to bool.
(maybe_add_script_text): Change return type and
loaded parameter to bool.
(source_script_file): Adjust.
(execute_script_contents): Adjust.
Change-Id: I59ab5862796fa7d154721b56e2ff8612ad5d734b
