[PATCH] RFC: use C11 alignas instead of GCC attribute aligned

[Mattias Rönnblom]

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 at lysator.liu.se]
>>>>>> Sent: Saturday, 27 January 2024 20.15
>>>>>>
>>>>>> On 2024-01-26 11:18, Morten Brørup wrote:
>>>>>>>> From: Mattias Rönnblom [mailto:hofors at lysator.liu.se]
>>>>>>>> Sent: Friday, 26 January 2024 11.05
>>>>>>>>
>>>>>>>> On 2024-01-25 23:53, Morten Brørup wrote:
>>>>>>>>>> From: Tyler Retzlaff [mailto:roretzla at 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
>>>