On 4/4/25 08:46, Michal Prívozník wrote:
> On 4/3/25 18:28, Roman Bogorodskiy wrote:
>> Michal Prívozník wrote:
>>
>>> On 4/2/25 19:24, Roman Bogorodskiy wrote:
>>>> The 'plain' optimization type also triggers the clang stack frame size
>>>> issues, so increase limit for it as well.
>>>>
>>>> Signed-off-by: Roman Bogorodskiy <bogorods...@gmail.com>
>>>> ---
>>>> meson.build | 2 +-
>>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/meson.build b/meson.build
>>>> index 56823ca25b..0a402a19a2 100644
>>>> --- a/meson.build
>>>> +++ b/meson.build
>>>> @@ -259,7 +259,7 @@ alloc_max = run_command(
>>>> stack_frame_size = 2048
>>>>
>>>> # clang without optimization enlarges stack frames in certain corner cases
>>>> -if cc.get_id() == 'clang' and get_option('optimization') == '0'
>>>> +if cc.get_id() == 'clang' and get_option('optimization') in ['plain', '0']
>>>> stack_frame_size = 4096
>>>> endif
>>>>
>>>
>>> Funny, with clang I hit this issue for all possible values of
>>> --optimization {plain,0,g,1,2,3,s}.
>>
>> That's interesting indeed.
>>
>> Currently, "plain" is the only value that triggers the issue for me:
>>
>> Problematic file is the following:
>>
>> ../src/remote/remote_driver.c:790:1: error: stack frame size (2344) exceeds
>> limit (2048) in 'doRemoteOpen' [-Werror,-Wframe-larger-than]
>> 790 | doRemoteOpen(virConnectPtr conn,
>> | ^
>> 1 error generated.
>>
>> Clang version is 19.1.7.
>
> Same here:
>
> clang version 19.1.7
> Target: x86_64-pc-linux-gnu
>
> except I'm running Linux and I assume you're running *BSD. Given what
> also Dan said, are you willing to post a v2 where the condition is
> simplified to just 'clang' and no get_option('optimization')?
>
> Actually, give me a minute or to - maybe there's something I can do
> about doRemoteOpen() so that its stack isn't that huge.
Alight, I think I know what's going on and why we're getting 2300+ bytes
long stack even with just a few variables declared.
Shortly: it's because of glib.
Longer version:
glib declares plenty of helper variables (mostly for type checking). For
instance:
g_clear_pointer() is declared as:
#define g_clear_pointer(pp, destroy) \
G_STMT_START \
{ \
G_STATIC_ASSERT (sizeof *(pp) == sizeof (gpointer)); \
glib_typeof ((pp)) _pp = (pp); \
glib_typeof (*(pp)) _ptr = *_pp; \
*_pp = NULL; \
if (_ptr) \
(destroy) (_ptr); \
} \
G_STMT_END \
And our VIR_FREE() is defined to be g_clear_pointer().
Then, g_strdup() is defined to be g_strdup_inline(), which: a) is
declared as always inline, and b) declares two additional variables:
G_ALWAYS_INLINE static inline char *
g_strdup_inline (const char *str)
{
if (__builtin_constant_p (!str) && !str)
return NULL;
if (__builtin_constant_p (!!str) && !!str && __builtin_constant_p
(strlen (str)))
{
const size_t len = strlen (str) + 1;
char *dup_str = (char *) g_malloc (len);
return (char *) memcpy (dup_str, str, len);
}
return g_strdup (str);
}
And finally, for some reason, attribute cleanup also increases stack
size (observed by playing with godbolt.org) - but this is for both gcc
and clang and isn't glib related. Honestly, I do not understand why
attribute cleanup would need to increase stack size, but whatever. I
think now I have all the ingredients needed to break the function down
into smaller pieces. So hopefully, in the end, this whole workaround
might be dropped.
NB: there's still vboxSnapshotRedefine() which has even bigger stack:
../src/vbox/vbox_common.c:4557:1: error: stack frame size (2824) exceeds
limit (2048) in 'vboxSnapshotRedefine' [-Werror,-Wframe-larger-than]
but I think (wishfully) the same strategy can be applied there.
Michal
