Re: Heads up: gcc miscompiling initramfs zlib decompression code at -O3

[Vineet Gupta]

On 4/30/21 3:06 PM, Linus Torvalds wrote:
> On Fri, Apr 30, 2021 at 1:46 PM Vineet Gupta <Vineet.Gupta1@xxxxxxxxxxxx> wrote:
>>
>> I've hit a mainline gcc 10.2 (also gcc 9.3) bug which triggers at -O3
>> causing wrong codegen.
> 
> I'd be more than happy to just disable CC_OPTIMIZE_FOR_PERFORMANCE_O3 entirely.
> 
> The advantages are very questionable - with a lot of the optimizations
> at O3 being about loops, something which the kernel to a close
> approximation doesn't have.
> 
> Most kernel loops are "count on one hand" iterations, and loop
> optimizations generally just make things worse.
> 
> And we've had problems with -O3 before, because not only are the
> optimizations a bit esoteric, they are often relatively untested. If
> you look around at various projects (outside the kernel), -O2 is
> generally the "default".

I agree that -O2 is default, but we've had -O3 default for ARC kernel 
forever, since last decade seriously. The reason I turned it on back 
then was upside of 10% performance improvement on select LMBench numbers 
on hardware at the time which for a rookie kernel hacker was yay momemt. 
I can revisit this and see if that is still true.

> And that's entirely ignoring the gcc history - where -O3 has often
> been very buggy indeed. It's gotten much better, but I just don't see
> the upside of using -O3.
> 
> In fact, it looks like we already have that
> 
>          depends on ARC
> 
> for -O3, exactly because nobody really wants to use this.

Either that or that people are not brave enough ;-) Perhaps gcc folks 
would like me to retain this as a testing ground if nothing else.

> So this bug seems to be entirely ARC-specific, in that only ARC can
> use -O3 for the kernel already.

kid in me complaining "that's not fair !"

-Vineet