init_fixed_point_type currently takes an objfile and creates its own
type allocator. However, for a later patch it is more convenient if
this function accepts a type allocator. This patch makes this change.
This replaces some casts to 'tracepoint *' with checked_static_cast.
Some functions are changed to accept a 'tracepoint *' now, for better
type safety.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
This replaces some casts to 'watchpoint *' with checked_static_cast.
In one spot, an unnecessary block is also removed.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Make sure to unlink the related breakpoint when the watchpoint instance
is deleted. This prevents having a wp-related breakpoint that is
linked to a NULL watchpoint (e.g. the watchpoint instance is being
deleted when the 'watch' command fails). With the below scenario,
having such a left out breakpoint will lead to a GDB hang, and this
is due to an infinite loop when deleting all inferior breakpoints.
Scenario:
(gdb) set can-use-hw-watchpoints 0
(gdb) awatch <SCOPE VAR>
Can't set read/access watchpoint when hardware watchpoints are disabled.
(gdb) rwatch <SCOPE VAR>
Can't set read/access watchpoint when hardware watchpoints are disabled.
(gdb) <continue the program until the end>
>> HANG <<
Signed-off-by: Mohamed Bouhaouel <mohamed.bouhaouel@intel.com>
Reviewed-by: Bruno Larsen <blarsen@redhat.com>
After the series that added this command was pushed, Pedro mentioned
that the news description could easily be misinterpreted, as well as
some code and test improvements that should be made.
While fixing the test, I realized that code repetition wasn't
happening as it should, so I took care of that too.
Approved-By: Andrew Burgess <aburgess@redhat.com>
Reviewed-By: Eli Zaretskii <eliz@gnu.org>
This patch changes class symbol_search to store a block_enum rather
than an int.
Regression tested on x86-64 Fedora 38.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
I stumbled across an ancient FIXME comment that was easy to fix --
val_prettyformat does not need to be in defs.h, and is easily moved to
valprint.h, where (despite what the comment says) it belongs.
Tested by rebuilding.
Add a unit test which checks that write_gdb_index_1 will throw
an error when the size of the file would exceed the maximum value
capable of being represented by 'offset_type'.
The unit test fails on 32-bit systems due to wrapping overflow. Fix this by
changing the type of total_len in write_gdbindex_1 from size_t to uint64_t.
Tested on x86_64-linux.
Co-Authored-By: Kevin Buettner <kevinb@redhat.com>
Approved-by: Kevin Buettner <kevinb@redhat.com>
I don't think these are useful nowadays, since we now expect all code to
be -Werror clean (it's the default in development branches).
Change-Id: I8c3b86c70d683bd41344d27add0ac2627a474d20
Approved-By: Tom Tromey <tom@tromey.com>
The amd-dbgapi library exposes a setting called "memory precision" for
AMD GPUs [1]. Here's a copy of the description of the setting:
The AMD GPU can overlap the execution of memory instructions with other
instructions. This can result in a wave stopping due to a memory violation
or hardware data watchpoint hit with a program counter beyond the
instruction that caused the wave to stop.
Some architectures allow the hardware to be configured to always wait for
memory operations to complete before continuing. This will result in the
wave stopping at the instruction immediately after the one that caused the
stop event. Enabling this mode can make execution of waves significantly
slower.
Expose this option through a new "amdgpu precise-memory" setting.
The precise memory setting is per inferior. The setting is transferred
from one inferior to another when using the clone-inferior command, or
when a new inferior is created following an exec or a fork.
It can be set before starting the inferior, in which case GDB will
attempt to apply what the user wants when attaching amd-dbgapi. If the
user has requested to enable precise memory, but it can't be enabled
(not all hardware supports it), GDB prints a warning.
If precise memory is disabled, GDB prints a warning when hitting a
memory exception (translated into GDB_SIGNAL_SEGV or GDB_SIGNAL_BUS),
saying that the stop location may not be precise.
Note that the precise memory setting also affects memory watchpoint
reporting, but the watchpoint support for AMD GPUs hasn't been
upstreamed to GDB yet. When we do upstream watchpoint support, GDB will
produce a similar warning message when stopping due to a watchpoint if
precise memory is disabled.
Add a handful of tests. Add a util proc
"hip_devices_support_precise_memory", which indicates if all devices
used for testing support that feature.
[1] 687374258a/include/amd-dbgapi.h.in (L6300-L6317)
Change-Id: Ife1a99c0e960513da375ced8f8afaf8e47a61b3f
Approved-By: Lancelot Six <lancelot.six@amd.com>
ROCm / HIP tests should only run on Linux for now, existing gdb.rocm
tests miss such a check. Add an "istarget linux" check in
allow_hipcc_tests.
Change-Id: I71f69e510a754f2fdadc32de53b923ebb9835ab5
Approved-By: Lancelot Six <lancelot.six@amd.com>
The following patch makes the amdgpu port transfer a property from the
original inferior to the new inferior when using the clone-inferior
command. Add the inferior_cloned observable to help with this.
Change-Id: Id845a799813ec49b1b7b2fcb97b07d0a1e5e2631
Approved-By: Tom Tromey <tom@tromey.com>
A user pointed out that the build failed with GCC 4.8. The problem
was that the form used by the std::hash specialization of ptid_t was
not accepted. This patch rewrites this code into a form that is
acceptable to the older compiler.
Approved-By: Simon Marchi <simon.marchi@efficios.com>
After commit 9d7d58e726, the amdgpu target started printing
"thread exited" messages when pruning waves that had terminated.
...
[AMDGPU Wave ?:?:?:2045 (?,?,?)/? exited]
[AMDGPU Wave ?:?:?:2046 (?,?,?)/? exited]
[AMDGPU Wave ?:?:?:2047 (?,?,?)/? exited]
[AMDGPU Wave ?:?:?:2048 (?,?,?)/? exited]
...
The issue was that before commit 9d7d58e726, delete_thread was silent
by default due to a bug that the commit fixed.
Replaced the amdgpu target call to delete_thread with a call to
delete_thread_silent.
Change-Id: Ie5d5a4c5be851f092d2315b2afa6a36a30a05245
Approved-By: Simon Marchi <simon.marchi@efficios.com>
Lancelot has accepted to take the role of maintainer for the AMD GPU
port. The AMD GPU port (amdgpu as I've written in the MAINTAINERS file)
is an umbrella term for everything needed to make this work: the amdgcn
arch, the amd-dbgapi target, solib-rocm, etc.
Thanks for accepting the role, and congratulations!
Change-Id: I4c898042fda49b45dcb0d54ca94731bb57287f71
This renames split_style::DOT, to avoid name clashes when building gdb
with an old version of Bison (2.3, the version available on macOS).
In particular the error looks like:
./split-name.h:34:3: error: expected identifier
DOT,
^
m2-exp.c:163:13: note: expanded from macro 'DOT'
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=30286
The R_ARC_TLS_LE_32 is defined as S + A + TLS_TBSS - TLS_REL, where
- S is the base address of the symbol in the memory
- A is the symbol addendum
- TLS_TBSS is the TLS Translation Control Block size (aligned)
- TLS_REL is the base of the TLS section
Given the next code snip:
__thread int data_var = 12;
__attribute__((__aligned__(128))) __thread int data_var_128 = 128;
__thread int bss_var;
__attribute__((__aligned__(256))) __thread int bss_var_256;
int __start(void)
{
return data_var + data_var_128 + bss_var + bss_var_256;
}
The current code returns different TLS_TBSS values for .tdata and
.tbss. This patch fixes this by using the linker provided tls_sec.
bfd/
* elf32-arc.c (TLS_REL): Clean up.
(TLS_TBSS): Use tls_sec alignment.
(arc_do_relocation): Check if we have valid tls_sec.
Signed-off-by: Claudiu Zissulescu <claziss@gmail.com>
While looking at how gdb::observers::new_objfile was used, I found
some code in symbol_file_add_with_addrs that I thought could be
improved.
Instead of:
if (condition)
{
...
return;
}
...
return;
Where some parts of '...' identical between the two branches. I think
it would be nicer if the duplication is removed, and we just use:
if (!condition)
...
to guard the one statement that should only happen when the condition
is not true.
There is one change in this commit though that is (possibly)
significant, there is a call to bfd_cache_close_all() that was only
present in the second block. After this commit we now call that
function for both paths.
The call to bfd_cache_close_all was added in commit:
commit ce7d45220e
Date: Fri Jul 30 12:05:45 2004 +0000
with the purpose of ensuring that GDB doesn't hold the BFDs open
unnecessarily, thus preventing the files from being updated on some
hosts (e.g. Win32).
In the early exit case we previously didn't call bfd_cache_close_all,
with the result that GDB would continue to hold open some BFD objects
longer than needed.
After this commit, but calling bfd_cache_close_all for both paths this
problem is solved.
I'm not sure how this change could be tested, I don't believe there's
any GDB (maintenance) command that displays the BFD cache contents, so
we can't check the cache contents easily. Ideas are welcome though.
Approved-By: Tom Tromey <tom@tromey.com>
The md_pre_output_hook creating fixup is asynchronous, causing relocs
may be out of order in .eh_frame. Define GAS_SORT_RELOCS so that reorder
relocs when write_relocs.
Reported-by: Rui Ueyama <rui314@gmail.com>
Now that CpuLM is used solely in cpu_arch_flags and cpu_arch[] while
Cpu64 is solely used in insn templates, they no longer need to be
treated different from other "ordinary" flags; the only "unusual" one
left if CpuNo64. Fold both, leaving just Cpu64.
A total four places exists where we set the two bits from flag_code, but
these values are never used. The two bits are evaluated only when coming
from insn templates.
Drop these assignments. Also make obvious that cpu_flags_check_cpu64()
is only ever used against insn templates.
While update_code_flag() checks for LM / i386, set_cpu_arch() so far
didn't, allowing e.g. 64-bit code to be emitted after ".arch generic32".
Oddly enough a few of our testcases actually exhibit bad behavior (and
hence need minor adjustments).
As mentioned in commit 3f5bbc3e20, some
compilers such as clang don't add debug information about stderr by
default, leaving it to external debug packages.
This commit adds a way to check if GDB has access to stderr information
when in MI mode, and uses this new mechanism to skip the related section
of the test gdb.mi/mi-dprintf.exp. It also fixes an incorrect name for a
test in that file.
Co-Authored-By: Andrew Burgess <aburgess@redhat.com>
Approved-By: Kevin Buettner <kevinb@redhat.com>
The header in a .gdb_index section uses 32-bit unsigned offsets to
refer to other areas of the section. Thus, there is a size limit of
2^32-1 which is currently unaccounted for by GDB's code for outputting
these sections.
At the moment, when GDB creates an overly large section, it will exit
abnormally due to an internal error, which is caused by a failed
assert in assert_file_size, which in turn is called from
write_gdbindex_1, both of which are in gdb/dwarf2/index-write.c.
This is what happens when that assert fails:
$ gdb -q -nx -iex 'set auto-load no' -iex 'set debuginfod enabled off' -ex file ./libgraph_tool_inference.so -ex "save gdb-index `pwd`/"
Reading symbols from ./libgraph_tool_inference.so...
No executable file now.
Discard symbol table from `libgraph_tool_inference.so'? (y or n) n
Not confirmed.
../../gdb/dwarf2/index-write.c:1069: internal-error: assert_file_size: Assertion `file_size == expected_size' failed.
A problem internal to GDB has been detected,
further debugging may prove unreliable.
----- Backtrace -----
0x55fddb4d78b0 gdb_internal_backtrace_1
../../gdb/bt-utils.c:122
0x55fddb4d78b0 _Z22gdb_internal_backtracev
../../gdb/bt-utils.c:168
0x55fddb98b5d4 internal_vproblem
../../gdb/utils.c:396
0x55fddb98b8de _Z15internal_verrorPKciS0_P13__va_list_tag
../../gdb/utils.c:476
0x55fddbb71654 _Z18internal_error_locPKciS0_z
../../gdbsupport/errors.cc:58
0x55fddb5a0f23 assert_file_size
../../gdb/dwarf2/index-write.c:1069
0x55fddb5a1ee0 assert_file_size
/usr/include/c++/13/bits/stl_iterator.h:1158
0x55fddb5a1ee0 write_gdbindex_1
../../gdb/dwarf2/index-write.c:1119
0x55fddb5a51be write_gdbindex
../../gdb/dwarf2/index-write.c:1273
[...]
---------------------
../../gdb/dwarf2/index-write.c:1069: internal-error: assert_file_size: Assertion `file_size == expected_size' failed.
This problem was encountered while building the python-graph-tool
package on Fedora. The Fedora bugzilla bug can be found here:
https://bugzilla.redhat.com/show_bug.cgi?id=1773651
This commit prevents the internal error from occurring by calling error()
when the file size exceeds 2^32-1.
Using a gdb built with this commit, I now see this behavior instead:
$ gdb -q -nx -iex 'set auto-load no' -iex 'set debuginfod enabled off' -ex file ./libgraph_tool_inference.so -ex "save gdb-index `pwd`/"
Reading symbols from ./libgraph_tool_inference.so...
No executable file now.
Discard symbol table from `/mesquite2/fedora-bugs/1773651/libgraph_tool_inference.so'? (y or n) n
Not confirmed.
Error while writing index for `/mesquite2/fedora-bugs/1773651/libgraph_tool_inference.so': gdb-index maximum file size of 4294967295 exceeded
(gdb)
I wish I could provide a test case, but due to the sizes of both the
input and output files, I think that testing resources would be
strained or exceeded in many environments.
My testing on Fedora 38 shows no regressions.
Approved-by: Tom Tromey <tom@tromey.com>
In this commit:
commit 48ac197b0c
Date: Fri Nov 19 10:12:44 2021 -0700
Handle multiple addresses in call_site_target
a buffer overflow bug was introduced when the following code was
added:
CORE_ADDR *saved = XOBNEWVAR (&objfile->objfile_obstack, CORE_ADDR,
addresses.size ());
std::copy (addresses.begin (), addresses.end (), saved);
The definition of XOBNEWVAR is (from libiberty.h):
#define XOBNEWVAR(O, T, S) ((T *) obstack_alloc ((O), (S)))
So 'saved' is going to point to addresses.size () bytes of memory,
however, the std::copy will write addresses.size () number of
CORE_ADDR sized entries to the address pointed to by 'saved', this is
going to result in memory corruption.
The mistake is that we should have used XOBNEWVEC, which allocates a
vector of entries, the definition of XOBNEWVEC is:
#define XOBNEWVEC(O, T, N) \
((T *) obstack_alloc ((O), sizeof (T) * (N)))
Which means we will have set aside enough space to create a copy of
the contents of the addresses vector.
I'm not sure how to create a test for this problem, this issue cropped
up when debugging a particular i686 built binary, which just happened
to trigger a glibc assertion (likely due to random memory corruption),
debugging the same binary built for x86-64 appeared to work just fine.
Using valgrind on the failing GDB binary pointed straight to the cause
of the problem, and with this patch in place there are no longer
valgrind errors in this area.
If anyone has ideas for a test I'm happy to work on something.
Co-Authored-By: Keith Seitz <keiths@redhat.com>
Approved-By: Tom Tromey <tom@tromey.com>
I've been running the test-suite on an i686-linux laptop with 1GB of memory,
and 1 GB of swap, and noticed problems after running gdb.base/huge.exp: gdb
not being able to spawn for a large number of test-cases afterwards.
So I investigated the memory usage, on my usual x86_64-linux development
platform.
The test-case is compiled with -DCRASH_GDB=2097152, so this:
...
static int a[CRASH_GDB], b[CRASH_GDB];
...
with sizeof (int) == 4 represents two arrays of 8MB each.
Say we add a loop around the "print a" command and print space usage
statistics:
...
gdb_test "maint set per-command space on"
for {set i 0} {$i < 100} {incr i} {
gdb_test "print a"
}
...
This gets us:
...
(gdb) print a^M
$1 = {0 <repeats 2097152 times>}^M
Space used: 478248960 (+469356544 for this command)^M
(gdb) print a^M
$2 = {0 <repeats 2097152 times>}^M
Space used: 486629376 (+8380416 for this command)^M
(gdb) print a^M
$3 = {0 <repeats 2097152 times>}^M
Space used: 495009792 (+8380416 for this command)^M
...
(gdb) print a^M
$100 = {0 <repeats 2097152 times>}^M
Space used: 1308721152 (+8380416 for this command)^M
...
In other words, we start out at 8MB, and the first print costs us about 469MB,
and subsequent prints 8MB, which accumulates to 1.3 GB usage. [ On the
i686-linux laptop, the first print costs us 335MB. ]
The subsequent 8MBs are consistent with the values being saved into the value
history, but the usage for the initial print seems somewhat excessive.
There is a PR open about needing sparse representation of large arrays
(PR8819), but this memory usage points to an independent problem.
The function value_print_array_elements contains a scoped_value_mark to free
allocated values in the outer loop, but it doesn't prevent the inner loop from
allocating a lot of values.
Fix this by adding a scoped_value_mark in the inner loop, after which we have:
...
(gdb) print a^M
$1 = {0 <repeats 2097152 times>}^M
Space used: 8892416 (+0 for this command)^M
(gdb) print a^M
$2 = {0 <repeats 2097152 times>}^M
Space used: 8892416 (+0 for this command)^M
(gdb) print a^M
$3 = {0 <repeats 2097152 times>}^M
Space used: 8892416 (+0 for this command)^M
...
(gdb) print a^M
$100 = {0 <repeats 2097152 times>}^M
Space used: 8892416 (+0 for this command)^M
...
Note that the +0 here just means that the mallocs did not trigger an sbrk.
This is dependent on malloc (which can use either mmap or sbrk or some
pre-allocated memory) and will likely vary between different tunings, versions
and implementations, so this does not give us a reliable way detect the
problem in a minimal way.
A more reliable way of detecting the problem is:
...
void
value_free_to_mark (const struct value *mark)
{
+ size_t before = all_values.size ();
auto iter = std::find (all_values.begin (), all_values.end (), mark);
if (iter == all_values.end ())
all_values.clear ();
else
all_values.erase (iter + 1, all_values.end ());
+ size_t after = all_values.size ();
+ if (before - after >= 1024)
+ fprintf (stderr, "value_free_to_mark freed %zu items\n", before - after);
...
which without the fix tells us:
...
+print a
value_free_to_mark freed 2097152 items
$1 = {0 <repeats 2097152 times>}
...
Fix a similar problem for Fortran:
...
+print array1
value_free_to_mark freed 4194303 items
$1 = (0, <repeats 2097152 times>)
...
in fortran_array_printer_impl::process_element.
The problem also exists for Ada:
...
+print Arr
value_free_to_mark freed 2097152 items
$1 = (0 <repeats 2097152 times>)
...
but is fixed by the fix for C.
Add Fortran and Ada variants of the test-case. The *.exp files are similar
enough to the original to keep the copyright years range.
While writing the Fortran test-case, I ran into needing an additional print
setting to print the entire array in repeat form, filed as PR exp/30817.
I managed to apply the compilation loop for the Ada variant as well, but with
a cumbersome repetition style. I noticed no other test-case uses gnateD, so
perhaps there's a better way of implementing this.
The regression test included in the patch is formulated in its weakest
form, to avoid false positive FAILs, which also means that smaller regressions
may not get detected.
Tested on x86_64-linux.
Approved-By: Tom Tromey <tom@tromey.com>
Rewrite test-case gdb.base/huge.exp:
- use build_executable rather than gdb_compile,
- use save_vars,
- factor out hardcoded loop limits min and max,
- handle compilation failure using require, and
- avoid using . in regexp to match $, {} and <>.
Tested on x86_64-linux.
Approved-By: Tom Tromey <tom@tromey.com>
Looking at the VEX and EVEX forms of vcvtneps2bf16 I noticed that
operand purpose isn't properly reflected in Vxy's definition. Rename
"dst" to "src", thus bringing things in line with Exy.
Recognize "/<number>" suffixes on both -march=+avx10.1 and the
corresponding .arch directive, setting an upper bound on the vector size
that insns may use. Such a restriction can be reset by setting a new base
architecture, by using a suffix-less form, by disabling AVX10, or by
enabling any other VEX/EVEX-based vector extension.
While for most insns we can suppress their use with too wide operands
via registers becoming unavailable (or in Intel syntax memory operand
size specifiers not being recognized), mask register insns have to have
their minimum required vector size specified in a new attribute. (Of
course this new attribute could also be used on other insns.)
Note that .insn continues to be permitted to emit EVEX{512,256} (and
VEX256 ones) encodings regardless of vector size restrictions in place.
Of course these can't be expressed using zmm (or ymm) operands then,
but need using the EVEX.512.* forms (broadcast forms may be usable right
now, but this may go away so shouldn't be relied upon). This is why no
assertions should be added to build_{e,}vex_prefix().
Since this is merely a re-branding of certain AVX512* features, there's
little code to be added.
The main aspect here are new testcases. In order to be able to re-use
some of the existing testcases, several of them need their start symbols
adjusted. Note that 256- and 128-bit tests want adding here, as these
need to work right away. Subsequently they'll gain vector length
constraints.
Since it was missing and is wanted here, also add an AVX512VL+VPOPCNTDQ
test.
These probably should have been put in place already anyway, but they're
very much wanted in order to then put AVX10.1 support on top. Note that
to avoid reverse dependencies towards SSE (just like we already do for
AVX and XOP), add_isa_dependencies() needs some further tweaking.
While there also address a related anomaly: Disabling AES but neither
AVX nor VAES (similarly for {,V}PCLMULQDQ) would better keep the 128-bit
VEX-encoded forms available. Note that for this the VAES insns are moved
past the AVX+AES ones, to avoid the property-11 test suddenly failing.
The test really is wrong, but let's not also make things inconsistent:
Without the movement, YMM use would be correctly recorded for the
128-bit forms simply because the first template already matches, as long
as VAES wasn't disabled. Yet it still wouldn't be if only AVX+AES were
enabled. Nor would behavior here then be the same as for VPCLMUL* insns.
Remove a warning about mi1. mi1 was removed in 975249ff4e ("Remove MI
version 1"). It is no longer possible to reach this warning, since
trying to use interpreter mi1 bails out before:
$ ./gdb -nx -q --data-directory=data-directory -i mi1
Interpreter `mi1' unrecognized
Change-Id: Ie43b21e01bca1407995150c729531a70ee662003
Approved-By: Tom Tromey <tom@tromey.com>
A user pointed out that if a DAP setBreakpoints request has a 'source'
field in a SourceBreakpoint object, then the gdb DAP implementation
will throw an exception.
While SourceBreakpoint does not allow 'source' in the spec, it seems
better to me to accept it. I don't think we should fully go down the
"Postel's Law" path -- after all, we have the type-checker -- but at
the same time, if we do send errors, they should be intentional and
not artifacts of the implementation.
Bug: https://sourceware.org/bugzilla/show_bug.cgi?id=30820
I see the following warning when building GDB on FreeBSD/amd64 with
Clang 14,
======================================================================
CXX mdebugread.o
mdebugread.c:1069:3: error: variable 'f' is uninitialized when used here [-Werror,-Wuninitialized]
f->set_loc_enumval (tsym.value);
^
mdebugread.c:836:17: note: initialize the variable 'f' to silence this warning
struct field *f;
^
= nullptr
======================================================================
after digging a little, I realized that we can not simply do what
Clang 14 says.
The root cause of this issue is that we lost the initialization of
the variable 'f' in this commit,
commit 2774f2dad5
Date: Thu Aug 31 09:37:44 2023 +0200
[gdb/symtab] Factor out type::{alloc_fields,copy_fields}
we have made these modifications,
---------------------------------------------------------------------
--- a/gdb/mdebugread.c
+++ b/gdb/mdebugread.c
@@ -1034,9 +1034,7 @@ parse_symbol (SYMR *sh, union aux_ext *ax, char *ext_sh, int bigend,
t->set_code (type_code);
t->set_length (sh->value);
- t->set_num_fields (nfields);
- f = ((struct field *) TYPE_ALLOC (t, nfields * sizeof (struct field)));
- t->set_fields (f);
+ t->alloc_fields (nfields, false);
---------------------------------------------------------------------
The problem is that the variable 'f' is used in the second half of
parse_symbol, that's why Clang complained.
To fix this issue we need to ensure that the varibale 'f' is
initialized. Calling the fields method is an obvious way to fix this
issue.
Tested on FreeBSD/amd64 by rebuilding.
Approved-By: Tom de Vries <tdevries@suse.de>
The gdb/testsuite/gdb.rocm/multi-inferior-gpu.cpp testcase contains a
call to execl which does not have NULL as a last argument. This is
an invalid use of execl. This patch fixes this oversight.
Change-Id: I03b60abe30468d71ba5089b240c6d00f9b8883b2
Approved-By: Tom Tromey <tom@tromey.com>
