[PATCH] D36562: [Bitfield] Make the bitfield a separate location if it has width of legal integer type and its bit offset is naturally aligned for the type

[Xinliang David Li via llvm-commits]

On Mon, Sep 4, 2017 at 1:57 AM, Chandler Carruth <chandlerc at gmail.com>
wrote:

> On Sun, Sep 3, 2017 at 10:41 PM Hal Finkel via llvm-commits <
> llvm-commits at lists.llvm.org> wrote:
>
>> Nevertheless, I think that you've convinced me that this is a least-bad
>> solution. I'll want a flag preserving the old behavior. Something like
>> -fwiden-bitfield-accesses (modulo bikeshedding).
>>
> Several different approaches have been discussed in this thread, I'm not
> sure what you mean about "least-bad solution"...
>
> I remain completely unconvinced that we should change the default
> behavior. At most, I'm not strongly opposed to adding an attribute that
> indicates "please try to use narrow loads for this bitfield member" and is
> an error if applied to a misaligned or non-byte-sized bitfield member. But
> I remain strongly opposed to changing the default behavior.
>

Having an option and make it off by default for now is a fine solution. We
can also use this option to collect more empirical data on a wider range of
tests and architectures with the community's help.


> We have one or two cases that regress and are easily addressed by source
> changes (either to not use bitfields or to use an attribute). I don't think
> that is cause to change the lowering Clang has used for years and
> potentially regress many other use cases.
>

Note that this is not about that particular cases. It is more about doing
the right thing (to make compiler and HW work together to drive the best
performance for our users).

thanks,

David
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170904/47fbd5ed/attachment.html>