Re: subtle side effect of commit a1c48bb160f836

[Vineet Gupta]

+CC Claudiu - ARC gcc guru

On Thursday 18 June 2015 02:25 PM, Geert Uytterhoeven wrote:
> Hi Vineet,
> 
> On Thu, Jun 18, 2015 at 10:45 AM, Vineet Gupta
> <Vineet.Gupta1@xxxxxxxxxxxx> wrote:
>> On Thursday 18 June 2015 01:43 PM, Michal Marek wrote:
>>>>> Alternatively, as we already have CONFIG_CC_OPTIMIZE_FOR_SIZE,
>>>>> a(nother) Kconfig option may make sense.
>>> We can also introduce some ARCH_CFLAGS that is appended near the end of
>>> the list, and have arc/Makefile add its -O3 there. But I'd like to why
>>> the -O3 needs to be there in first place.
>>
>> This is how historically ARC kernels have been built. We do track performance
>> results LMBench/hackbench... and going from -O3 to -O2 caused a sudden dip in some
>> of the numbers when we measured 3.18 (vs. 3.13)
>>
>>> Obviously, the kernel works
>>> with -O2, otherwise the regression would have been identified earlier.
>>
>> Its a performance thing - so yeah -O2 works, but -O3 works even better :-)
> 
> Did you see some numbers increase when going from -O3 to -O2?
> IIRC, -O3 enables more aggressive inlining, which can cause more L1 cache
> misses.

It sure does but smaller functions could cause more stack return mispredicts etc.
It all boils down to the micro-arch in the end and how gcc does arch specific
things under the hood of -O{2,s}.


> It might be worth trying CONFIG_CC_OPTIMIZE_FOR_SIZE=y...

Not for ARC. At -Os gcc is more worried about using short instructions (2 bytes),
and things like alignment of target branches, ld/st scheduling might not be as
optim as with -O2/O3. Some of the code density instructions have associated
pipeline stalls etc.

So last time (it's been a while though) when I ran benchmarks with -Os on ARC, it
was way off vs. -O2.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/