Re: [PATCH v5] x86: use builtins to read eflags

From: Linus Torvalds <torvalds@linux-foundation.org>
To: Bill Wendling <morbo@google.com>
Cc: Nick Desaulniers <ndesaulniers@google.com>,
	"H. Peter Anvin" <hpa@zytor.com>,
	 Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@redhat.com>, Borislav Petkov <bp@alien8.de>,
	 Dave Hansen <dave.hansen@linux.intel.com>,
	 "maintainer:X86 ARCHITECTURE (32-BIT AND 64-BIT)"
	<x86@kernel.org>, Nathan Chancellor <nathan@kernel.org>,
	Juergen Gross <jgross@suse.com>,
	 Peter Zijlstra <peterz@infradead.org>,
	Andy Lutomirski <luto@kernel.org>,
	llvm@lists.linux.dev,  LKML <linux-kernel@vger.kernel.org>,
	 linux-toolchains <linux-toolchains@vger.kernel.org>
Subject: Re: [PATCH v5] x86: use builtins to read eflags
Date: Thu, 17 Mar 2022 16:14:41 -0700	[thread overview]
Message-ID: <CAHk-=wgsYO2BAv+WaeHzP5GkeiWfdUW7tVQ5CG-aVRwANWMBsg@mail.gmail.com> (raw)
In-Reply-To: <CAHk-=wj=aK2T+h=rRJL4jDDTgV2U533E_duroCoVSUfkvYd0OQ@mail.gmail.com>

On Thu, Mar 17, 2022 at 3:51 PM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> I still think that from a sanity standpoint, it would be good to
> actually strengthen the semantics of "asm volatile" to literally act
> as - and be ordered with - volatile memory accesses.

Side note: the reason I personally would prefer these semantics is
that if we had a stronger definition of what "asm volatile" actually
means, we could get rid of some - but certainly not all - "memory"
clobbers.

And while I've been very positive about memory clobbers in this thread
- because they give us that serialization we so often need - they can
often be bad for code generation.

Quite often it's not actually a huge deal. Even with a memory clobber,
it's not like the compiler needs to flush any spills or local
variables that haven't had their address taken to memory.

So quite often a memory clobber is only going to affect instruction
ordering a bit, and that's usually exactly what you want.

But it certainly _can_ be quite noticeable, and while things like
"cli" really wants a memory clobber anyway because it really does want
to order with respect to memory accesses that the compiler does, other
operations wouldn't necessarily need it.

That "native_save_fl()" that does that "pushf ; pop %0" instruction,
for example, would be perfectly happy only being ordered wrt other asm
instructions (notably ordered wrt cli/sti).

So we could easily remove the "memory" clobber there, but only if we
had that other ordering constraint of "asm volatile" being ordered wrt
each other.

Right now we have memory clobbers just about everywhere. It's almost
the default, unless it's purely about some purely arithmetic
operation. I'm not sure how many of them we could remove, but I do
think that pushf/popf is _one_ such case.

So stronger semantics for "asm volatile" could potentially help
generate better code.

Of course, then the "I just don't want you to optimize this away"
crowd might be unhappy and find cases where it really hurts.

               Linus