RFA (libstdc++): PATCH to implement C++17 over-aligned new

[Jonathan Wakely]

On 16/09/16 11:56 +0100, Jonathan Wakely wrote:
>On 16/09/16 11:37 +0200, Marc Glisse wrote:
>>On Fri, 16 Sep 2016, Jonathan Wakely wrote:
>>
>>>On 16/09/16 09:04 +0200, Rainer Orth wrote:
>>>>Hi Jason,
>>>>
>>>>>OK, one more:
>>>>
>>>>this works just fine on both sparc-sun-solaris2.12 and
>>>>i386-pc-solaris2.12.
>>>>
>>>>Once Jonathan's patch to heed aligned_alloc's requirement on size being
>>>>a multiple of alignment is in, all is fine on Solaris.
>>>
>>>I've got a slightly different fix now.
>>>
>>>We only need to make the size a multiple of alignment for
>>>aligned_alloc, however for posix_memalign we need to ensure the
>>>alignment is a multiple of sizeof(void*).
>>>
>>>I'm testing this now (but only on x86_64 GNU/Linux where it wasn't
>>>failing anyway).
>>
>>+  // The value of alignment shall be a power of two multiple of sizeof(void *).
>>+  if (al < sizeof(void*))
>>+    al = sizeof(void*);
>>
>>The code doesn't exactly match the comment. I can't find the 
>>precondition in the standard that says operator new can only be 
>>called on a power of 2... (maybe we can add it if it is really 
>>missing?)
>
>[basic.align] says "Every alignment value shall be a non-negative
>integral power of two." So asking operator new for any other value
>doesn't make sense, but I can't find a restriction on doing so.
>
>I was assuming we only need to ensure it's possible to use valid
>alignments such as align_val_t(2) which are not valid arguments to
>posix_memalign. For other values such as align_val_t(15) I was
>assuming it's OK for posix_memalign to fail, so we throw bad_alloc.
>
>If that's not the case then we need to round up all alignments that
>aren't power of two multiples of sizeof(void*). I'd like to avoid
>that.
>
>>>Would using __builtin_expect (sz == 0, false) make sense?  Surely it's
>>>rare to try to allocate zero bytes.
>>
>>https://gcc.gnu.org/ml/libstdc++/2014-03/msg00001.html
>>
>>gcc already guesses that a test like sz == 0 is usually false (not 
>>with as large a probability as if you use __builtin_expect, but 
>>enough that the generated code is unlikely to differ). But adding 
>>__builtin_expect cannot hurt...
>>
>>Is the division (by a non-constant denominator) really necessary?
>
>Probably not, but I've asked the committee for clarification what this
>function should do when called with an invalid alignment.
>
>>Since align has to be a power of 2, x % align should be the same as 
>>x & (align - 1), for instance.
>
>Thanks, if it's UB to call it with alignments that aren't a power of
>two then we can do that.

I've committed the patch now, to fix the failures for Solaris. I'll
revisit it when I get clarification from the committee about invalid
alignment arguments.