Re: -falign-functions=16 for i386/amd64

[Ryota Ozaki <ozaki-r%netbsd.org@localhost>]

On Fri, Sep 2, 2016 at 4:40 AM, Joerg Sonnenberger <joerg%bec.de@localhost> wrote:
> On Thu, Sep 01, 2016 at 10:00:52PM +0900, Ryota Ozaki wrote:
>> On Thu, Sep 1, 2016 at 6:45 PM, Joerg Sonnenberger <joerg%bec.de@localhost> wrote:
>> > On Thu, Sep 01, 2016 at 03:46:15PM +0900, Ryota Ozaki wrote:
>> >> http://www.netbsd.org/~ozaki-r/align-functions-16.diff
>> >>
>> >> The patch adds the option to sys/arch/amd64/conf/Makefile.amd64.
>> >> Is it a feasible place to add?
>> >
>> > There are two small issues I have with this patch:
>> > (1) I think it should be restricted to GCC with an appropiate comment of
>> > what this is a workaround for. Clang seems to behave a lot more sensible
>> > out of the box. If there are CPU models with a different base alignment
>> > and the user asked for one of them as optimisation target, it should be
>> > honored IMO.
>> > (2) This should not touch CFLAGS, but COPTS.
>>
>> Okay, I see. How about the following patch?
>> (nonaka@ helped improving Makefile options.)
>>
>> http://www.netbsd.org/~ozaki-r/align-functions-16.v2.diff
>
> Almost. You shouldn't need the whole if-block.

The if-block intends to add the option even when a kernel config
(like GENERIC) has own COPTS.

Well, should we diereclty add the option to COPTS in GENERIC instead?
Or we may be able to get rid of (or comment-out) COPTS from GENERIC
because it's the same as DEFCOPTS.

>  I don't understand the
> bsd.own.mk reference, it doesn't contain -mtune=nocona here?

I meant GCC_CONFIG_TUNE.x86_64=nocona line that makes gcc use
nocona as the default value of -mtune.

Okay, revised the comment like this:

-# By default, our gcc uses -mtune=nocona for compiling the kernels
-# (see share/mk/bsd.own.mk). With -mtune=nocona, gcc doesn't align
+# Our gcc is built to use nocona as the default value of -mtune
+# (see GCC_CONFIG_TUNE in share/mk/bsd.own.mk). With -mtune=nocona,
+# gcc doesn't align (...)

Thanks,
  ozaki-r