Re: [lttng-dev] [PATCH 02/11] urcu/uatomic: Use atomic builtins if configured

On 6/22/23 14:32, Paul E. McKenney wrote:
> On Thu, Jun 22, 2023 at 11:55:55AM -0400, Mathieu Desnoyers wrote:
>> On 6/21/23 19:19, Paul E. McKenney wrote:
>> [...]
>>>> diff --git a/include/urcu/uatomic/builtins-generic.h b/include/urcu/uatomic/builtins-generic.h
>>>> new file mode 100644
>>>> index 0000000..8e6a9b5
>>>> --- /dev/null
>>>> +++ b/include/urcu/uatomic/builtins-generic.h
>>>> @@ -0,0 +1,85 @@
>>>> +/*
>>>> + * urcu/uatomic/builtins-generic.h
>>>> + *
>>>> + * Copyright (c) 2023 Olivier Dion <[email protected]>
>>>> + *
>>>> + * This library is free software; you can redistribute it and/or
>>>> + * modify it under the terms of the GNU Lesser General Public
>>>> + * License as published by the Free Software Foundation; either
>>>> + * version 2.1 of the License, or (at your option) any later version.
>>>> + *
>>>> + * This library is distributed in the hope that it will be useful,
>>>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>>>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
>>>> + * Lesser General Public License for more details.
>>>> + *
>>>> + * You should have received a copy of the GNU Lesser General Public
>>>> + * License along with this library; if not, write to the Free Software
>>>> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
>>>> + */
>>>> +
>>>> +#ifndef _URCU_UATOMIC_BUILTINS_GENERIC_H
>>>> +#define _URCU_UATOMIC_BUILTINS_GENERIC_H
>>>> +
>>>> +#include <urcu/system.h>
>>>> +
>>>> +#define uatomic_set(addr, v) __atomic_store_n(addr, v, __ATOMIC_RELAXED)
>>>> +
>>>> +#define uatomic_read(addr) __atomic_load_n(addr, __ATOMIC_RELAXED)
>>>
>>> Does this lose the volatile semantics that the old-style definitions
>>> had?
>>>
>>
>> Yes.
>>
>> [...]
>>
>>>> +++ b/include/urcu/uatomic/builtins-x86.h
>>>> @@ -0,0 +1,85 @@
>>>> +/*
>>>> + * urcu/uatomic/builtins-x86.h
>>>> + *
>>>> + * Copyright (c) 2023 Olivier Dion <[email protected]>
>>>> + *
>>>> + * This library is free software; you can redistribute it and/or
>>>> + * modify it under the terms of the GNU Lesser General Public
>>>> + * License as published by the Free Software Foundation; either
>>>> + * version 2.1 of the License, or (at your option) any later version.
>>>> + *
>>>> + * This library is distributed in the hope that it will be useful,
>>>> + * but WITHOUT ANY WARRANTY; without even the implied warranty of
>>>> + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
>>>> + * Lesser General Public License for more details.
>>>> + *
>>>> + * You should have received a copy of the GNU Lesser General Public
>>>> + * License along with this library; if not, write to the Free Software
>>>> + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
>>>> + */
>>>> +
>>>> +#ifndef _URCU_UATOMIC_BUILTINS_X86_H
>>>> +#define _URCU_UATOMIC_BUILTINS_X86_H
>>>> +
>>>> +#include <urcu/system.h>
>>>> +
>>>> +#define uatomic_set(addr, v) __atomic_store_n(addr, v, __ATOMIC_RELAXED)
>>>> +
>>>> +#define uatomic_read(addr) __atomic_load_n(addr, __ATOMIC_RELAXED)
>>>
>>> And same question here.
>>
>> Yes, this opens interesting questions:
>>
>> * what semantic do we want for uatomic_read/set ?
>>
>> * what semantic do we want for CMM_LOAD_SHARED/CMM_STORE_SHARED ?
>>
>> * do we want to allow load/store-shared to work on variables larger than a
>> word ? (e.g. on a uint64_t on a 32-bit architecture, or on a structure)
>>
>> * what are the guarantees of a volatile type ?
>>
>> * what are the guarantees of a load/store relaxed in C11 ?
>>
>> Does the delta between volatile and C11 relaxed guarantees matter ?
>>
>> Is there an advantage to use C11 load/store relaxed over volatile ? Should
>> we combine both C11 load/store _and_ volatile ? Should we use
>> atomic_signal_fence instead ?
> 
> I suggest C11 volatile atomic load/store.  Load/store fusing is permitted
> for non-volatile atomic loads and stores, and such fusing can ruin your
> code's entire day.  ;-)

I'm OK with erring towards a safer approach, but just out of curiosity, 
do you have examples of compilers doing load or store fusion on C11 or 
C++11 relaxed atomics, or is it out of caution due to lack of explicit 
guarantees in the standards ?

Does this lack of guarantee about fusion also apply to other MO such as 
acquire, release and seq.cst. ?

Thanks,

Mathieu


-- 
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com

_______________________________________________
lttng-dev mailing list
[email protected]
https://lists.lttng.org/cgi-bin/mailman/listinfo/lttng-dev