Re: [libstdc++/65033] Give alignment info to libatomic

[Jonathan Wakely <jwakely@redhat.com>]

On 31/03/15 08:13 -0700, Richard Henderson wrote:
>On 03/31/2015 08:03 AM, Jonathan Wakely wrote:
>> On 31/03/15 07:54 -0700, Richard Henderson wrote:
>>> On 03/31/2015 06:41 AM, Jonathan Wakely wrote:
>>>> This is the best I've come up with, does anyone have any better ideas
>>>> than the #else branch to hardcode alignment of 16-byte types to 16?
>>>
>>> If there's no 16 byte type, are we convinced this matters?  I mean, there isn't
>>> a 16-byte atomic instruction for 32-bit x86 (or any other 32-bit cpu of which I
>>> am aware).  So we're forced to use a locking path anyway.
>>
>> The C front end gives struct S { char s[16]; } 16 byte alignment...
>
>Um, I'm pretty sure it doesn't.
>
>	struct S { char s[16]; };
>	int x = __alignof(struct S);
>
>	.type	x, @object
>	.size	x, 4
>x:
>	.long	1
>
>What you're interpreting as alignment for that struct is probably optional
>*additional* alignment from LOCAL_ALIGNMENT or LOCAL_DECL_ALIGNMENT or something.

Sorry for not being clear, I meant __alignof(_Atomic struct S) is 16.

>> And it matters most for the integral types, not arbitrary structs.
>>
>>> And if we do want the alignment, do we stop pretending with all the sizeof's
>>> and alignof's and just use power-of-two size alignment for all of them, e.g.
>>>
>>>  min_align = ((size & (size - 1)) || size > 16 ? 0 : size)
>>
>> Yeah, I wondered about that too. Joseph indicated there are targets
>> where C gives alignof(_Atomic X) != sizeof(X), which is why the target
>> hook exists, but maybe we can just not worry about those targets for
>> now.
>
>Those targets have alignof < sizeof.  So in a way we'd probably be doing them a
>favor.  I know for instance that CRIS is in this category, where most data is
>all byte aligned, but atomics must be naturally aligned.

Aha, I wondered why CRIS overrides the atomic_align_for_mode hook when
it seemed to be giving them natural alignment anyway - I didn't
realise non-atomic types are only byte-aligned.