Re: [PATCH 1/6] riscv: Introduce use_alternative_{likely,unlikely}

From: Yury Norov <yury.norov@gmail.com>
To: Vivian Wang <wangruikang@iscas.ac.cn>
Cc: "Paul Walmsley" <paul.walmsley@sifive.com>,
	"Palmer Dabbelt" <palmer@dabbelt.com>,
	"Albert Ou" <aou@eecs.berkeley.edu>,
	"Alexandre Ghiti" <alex@ghiti.fr>,
	"Rasmus Villemoes" <linux@rasmusvillemoes.dk>,
	"Vivian Wang" <uwu@dram.page>,
	linux-riscv@lists.infradead.org, linux-kernel@vger.kernel.org,
	"Aydın Mercan" <aydin@mercan.dev>
Subject: Re: [PATCH 1/6] riscv: Introduce use_alternative_{likely,unlikely}
Date: Wed, 20 Aug 2025 11:43:00 -0400	[thread overview]
Message-ID: <aKXtBJ5Aqqmvdc1B@yury> (raw)
In-Reply-To: <544577b2-6810-4bef-b588-e1c662d5be13@iscas.ac.cn>

> > This 'unlikely' version is just an negation of 'likely' one, and it
> > looks like an attempt to save on one negation. On the other hand, the
> > function is __always_inline, which means that compiler should normally
> > take care of it. Can you prove with objdump that it really works as
> > intended? I mean that 
> >
> >         if (use_alternative_unlikely())
> >                 do_something();
> >
> > generates a better code than 
> >         
> >         if (!use_alternative_likely())
> >                 do_something();
> 
> use_alternative_likely() and use_alternative_unlikely() are not
> negations of each other and in fact should be functionally equivalent. I
> also briefly explained the difference in the comment, but the difference
> is which case is nop i.e. fallthrough, and which case requires a jump
> instruction. The likely case should get a "nop", and the unlikely case
> should get a "j %l[...]". This choice does work as intended [1].
> 
> I don't think it is possible to give both options to the compiler, so at
> least for now AIUI users have to pick one.
> 
> The same applies to __riscv_has_extension_{likely,unlikely}.
> 
> Vivian "dramforever" Wang
> 
> [1]: https://godbolt.org/z/v8zTEhzTx

I realize that likely and unlikely versions generate different code,
I'm just not convinced that

1. it works in real kernel as intended, not only in the godbold; and
2. has any measurable impact.

That's why I asked you to share objdump and possibly perf tests.

_______________________________________________
linux-riscv mailing list
linux-riscv@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-riscv