From: "Mattias Rönnblom" <hofors@lysator.liu.se>
To: Tyler Retzlaff <roretzla@linux.microsoft.com>
Cc: "Morten Brørup" <mb@smartsharesystems.com>,
dev@dpdk.org, "Mattias Rönnblom" <mattias.ronnblom@ericsson.com>,
"Anatoly Burakov" <anatoly.burakov@intel.com>,
"Bruce Richardson" <bruce.richardson@intel.com>,
"David Christensen" <drc@linux.vnet.ibm.com>,
"Harry van Haaren" <harry.van.haaren@intel.com>,
"Konstantin Ananyev" <konstantin.v.ananyev@yandex.ru>,
"Min Zhou" <zhoumin@loongson.cn>,
"Ruifeng Wang" <ruifeng.wang@arm.com>,
"Stanislaw Kardach" <kda@semihalf.com>,
thomas@monjalon.net
Subject: Re: [PATCH] RFC: use C11 alignas instead of GCC attribute aligned
Date: Wed, 31 Jan 2024 17:04:59 +0100 [thread overview]
Message-ID: <edf19856-69a2-47f5-b821-0151997d26ae@lysator.liu.se> (raw)
In-Reply-To: <20240130173928.GA2943@linuxonhyperv3.guj3yctzbm1etfxqx2vob5hsef.xx.internal.cloudapp.net>
On 2024-01-30 18:39, Tyler Retzlaff wrote:
> On Tue, Jan 30, 2024 at 09:08:21AM +0100, Mattias Rönnblom wrote:
>> On 2024-01-29 20:43, Tyler Retzlaff wrote:
>>> On Sun, Jan 28, 2024 at 11:00:31AM +0100, Mattias Rönnblom wrote:
>>>> On 2024-01-28 09:57, Morten Brørup wrote:
>>>>>> From: Mattias Rönnblom [mailto:hofors@lysator.liu.se]
>>>>>> Sent: Saturday, 27 January 2024 20.15
>>>>>>
>>>>>> On 2024-01-26 11:18, Morten Brørup wrote:
>>>>>>>> From: Mattias Rönnblom [mailto:hofors@lysator.liu.se]
>>>>>>>> Sent: Friday, 26 January 2024 11.05
>>>>>>>>
>>>>>>>> On 2024-01-25 23:53, Morten Brørup wrote:
>>>>>>>>>> From: Tyler Retzlaff [mailto:roretzla@linux.microsoft.com]
>>>>>>>>>> Sent: Thursday, 25 January 2024 19.37
>>>>>>>>>>
>>>>>>>>>> ping.
>>>>>>>>>>
>>>>>>>>>> Please review this thread if you have time, the main point of
>>>>>>>>>> discussion
>>>>>>>>>> I would like to receive consensus on the following questions.
>>>>>>>>>>
>>>>>>>>>> 1. Should we continue to expand common alignments behind an
>>>>>>>> __rte_macro
>>>>>>>>>>
>>>>>>>>>> i.e. what do we prefer to appear in code
>>>>>>>>>>
>>>>>>>>>> alignas(RTE_CACHE_LINE_MIN_SIZE)
>>>>>>>>>>
>>>>>>>>>> -- or --
>>>>>>>>>>
>>>>>>>>>> __rte_cache_aligned
>>>>>>>>>>
>>>>>>>>>> One of the benefits of dropping the macro is it provides a clear
>>>>>>>> visual
>>>>>>>>>> indicator that it is not placed in the same location or get
>>>>>> applied
>>>>>>>>>> to types as is done with __attribute__((__aligned__(n))).
>>>>>>>>>
>>>>>>>>> We don't want our own proprietary variant of something that already
>>>>>>>> exists in the C standard. Now that we have moved to C11, the __rte
>>>>>>>> alignment macros should be considered obsolete.
>>>>>>>>
>>>>>>>> Making so something cache-line aligned is not in C11.
>>>>>>>
>>>>>>> We are talking about the __rte_aligned() macro, not the cache
>>>>>> alignment macro.
>>>>>>>
>>>>>>
>>>>>> OK, in that case, what is the relevance of question 1 above?
>>>>>
>>>>> With this in mind, try re-reading Tyler's clarifications in this tread.
>>>>>
>>>>> Briefly: alignas() can be attached to variables and structure fields, but not to types (like __rte_aligned()), so to align a structure:
>>>>>
>>>>> struct foo {
>>>>> int alignas(64) bar; /* alignas(64) must be here */
>>>>> int baz;
>>>>> }; /* __rte_aligned(64) was here, but alignas(64) cannot be here. */
>>>>>
>>>>> So the question is: Do we want to eliminate the __rte_aligned() macro - which relies on compiler attributes - and migrate to using the C11 standard alignas()?
>>>>>
>>>>> I think yes; after updating to C11, the workaround for pre-C11 not offering alignment is obsolete, and its removal should be on the roadmap.
>>>>>
>>>>
>>>> OK, thanks for the explanation. Interesting limitation in the standard.
>>>>
>>>> If the construct the standard is offering is less effective (in this
>>>> case, less readable) and the non-standard-based option is possible
>>>> to implement on all compilers (i.e., on MSVC too), then we should
>>>> keep the custom option. Especially if it's already there, but also
>>>> in cases where it isn't.
>>>>
>>>> In fact, one could argue *everything* related to alignment should go
>>>> through something rte_, __rte_ or RTE_-prefixed. So, "int
>>>> RTE_ALIGNAS(64) bar;". Maybe that would be silly, but it would be
>>>> consistent with RTE_CACHE_ALIGNAS.
>>>>
>>>> I would worry more about allowing DPDK developers writing clean and
>>>> readable code, than very slightly lowering the bar for the fraction
>>>> of newcomers experienced with the latest and greatest from the C
>>>> standard, and *not* familiar with age-old GCC extensions.
>>>
>>> I’d just like to summarize where my understanding is at after reviewing
>>> this discussion and my downstream branch. But I also want to make it
>>> clear that we probably need to use both standard C and non-standard
>>> attribute/declspec for object and struct/union type alignment
>>> respectively.
>>>
>>> I've assumed we prefer avoiding per-compiler conditional expansion when
>>> possible through the use of standard C mechanisms. But there are
>>> instances when alignas is awkward.
>>>
>>> So I think the following is consistent with what Mattias is advocating
>>> sans any discussions related to actual naming of macros.
>>>
>>> We should have 2 macros, upon which others may be built to expand to
>>> well-known values for e.g. cache line size.
>>>
>>> RTE_ALIGNAS(n) object;
>>>
>>> * This macro is used to align C objects i.e. variable, array, struct/union
>>> fields etc.
>>> * Trivially expands to alignas(n) for all toolchains.
>>> * Placed in a location that both C and C++ translation units accept that
>>> is on the same line preceeding the object type.
>>> example:
>>> // RTE_ALIGNAS(n) object;
>>> RTE_ALIGNAS(16) char somearray[16];
>>>
>>> RTE_ALIGN_TYPE(n)
>>>
>>> * This macro is used to align struct/union types.
>>> * Conditionally expands to __declspec(align(n)) (msvc) and
>>> __attribute__((__aligned__(n))) (for all other toolchains)
>>> * Placed in a location that for all gcc,clang,msvc and both C and C++
>>> translation units accept.
>>> example:
>>> // {struct,union} RTE_ALIGN_TYPE(n) tag { ... };
>>> struct RTE_ALIGN_TYPE(64) sometype { ... };
>>>
>>
>> Sorry if I've missed some discussion on the list, but the current
>> pattern of putting __rte_aligned(X) at the end doesn't work with
>> MSVC, or why are we doing this? C11 purism doesn't seem like much of
>> a driving force.
>
> __rte_aligned(X) at the end doesn't work with MSVC __declspec(align(n))
>
>>
>> If one defined a macro as __declspec(align(X)) on MSVC and
>> __attribute__(__aligned__(X)) on other compilers, could it do the
>> work of both the above RTE_ALIGNAS() and RTE_ALIGN_TYPE()?
>>
>> <a> struct <b> { int a; } <c>;
>
> yes for struct/union. but only when placed at location you mark as <b>
> when compiling both C and C++ for all toolchains.
>
> maybe, for objects but ideally, we prefer alignas for consistent semantics
> defined by standard rather than accomodating potential implementation
> differences when conditionally expanding __aligned vs __declspec. as you
> have noted __declspec has limitations/variations when compared to
> __attribute__((__aligned__(n))).
>
>>
>> You would have to mandate the placement of such a __rte_aligned
>> plug-in replacement being at <b> rather than (the more intuitive?)
>> <a>, since clang doesn't like __attribute__s before the struct/union
>> keyword, correct?
>
> for struct/union there is a single placement accepted by all toolchains
> for both C and C++ and it is <b>.
>
>>
>> What about other <rte_common.h> __attribute__ wrappers like
>> __rte_packed; would they also need to change placement to make DPDK
>> work with MSVC?
>
> packing is a different problem that needs a separate RFC and discussion
> of it's own.
>
Seems like the same kind of problem with potentially the exact same
solution: mandate a new "__rte_xxx" placement and use MSVC __declspec.
Different RFC, yes, different discussion: not so sure.
>>
>>> I'm not picky about what the names actualy are if you have better
>>> suggestions i'm happy to adopt them.
>>>
>>> Thoughts? Comments?
>>>
>>> Appreciate the discussion this has been helpful.
>>>
>>> ty
>>>
next prev parent reply other threads:[~2024-01-31 16:05 UTC|newest]
Thread overview: 29+ messages / expand[flat|nested] mbox.gz Atom feed top
2023-11-15 17:39 Tyler Retzlaff
2023-11-15 17:39 ` [PATCH] eal: " Tyler Retzlaff
2023-11-15 18:13 ` Bruce Richardson
2023-11-15 18:27 ` Tyler Retzlaff
2023-11-15 20:08 ` Morten Brørup
2023-11-15 21:03 ` Tyler Retzlaff
2023-11-15 22:43 ` Stanisław Kardach
2023-11-16 10:12 ` Mattias Rönnblom
2024-01-25 18:37 ` [PATCH] RFC: " Tyler Retzlaff
2024-01-25 22:53 ` Morten Brørup
2024-01-25 23:31 ` Tyler Retzlaff
2024-01-26 10:05 ` Mattias Rönnblom
2024-01-26 10:18 ` Morten Brørup
2024-01-27 19:15 ` Mattias Rönnblom
2024-01-28 8:57 ` Morten Brørup
2024-01-28 10:00 ` Mattias Rönnblom
2024-01-29 19:43 ` Tyler Retzlaff
2024-01-30 8:08 ` Mattias Rönnblom
2024-01-30 17:39 ` Tyler Retzlaff
2024-01-30 17:59 ` Bruce Richardson
2024-01-30 18:01 ` Bruce Richardson
2024-01-30 18:04 ` Tyler Retzlaff
2024-01-30 18:18 ` Mattias Rönnblom
2024-01-31 16:04 ` Mattias Rönnblom [this message]
2024-01-30 8:09 ` Morten Brørup
2024-01-30 9:28 ` Mattias Rönnblom
2024-01-30 10:17 ` Morten Brørup
2024-01-30 13:00 ` Morten Brørup
2024-01-30 17:54 ` Tyler Retzlaff
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=edf19856-69a2-47f5-b821-0151997d26ae@lysator.liu.se \
--to=hofors@lysator.liu.se \
--cc=anatoly.burakov@intel.com \
--cc=bruce.richardson@intel.com \
--cc=dev@dpdk.org \
--cc=drc@linux.vnet.ibm.com \
--cc=harry.van.haaren@intel.com \
--cc=kda@semihalf.com \
--cc=konstantin.v.ananyev@yandex.ru \
--cc=mattias.ronnblom@ericsson.com \
--cc=mb@smartsharesystems.com \
--cc=roretzla@linux.microsoft.com \
--cc=ruifeng.wang@arm.com \
--cc=thomas@monjalon.net \
--cc=zhoumin@loongson.cn \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).