Re: [RFC PATCH] LKMM: Add ctrl_dep() macro for control dependency

From: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
To: Segher Boessenkool <segher@kernel.crashing.org>
Cc: Florian Weimer <fweimer@redhat.com>,
	Will Deacon <will@kernel.org>, paulmck <paulmck@kernel.org>,
	Peter Zijlstra <peterz@infradead.org>,
	linux-kernel <linux-kernel@vger.kernel.org>,
	Linus Torvalds <torvalds@linux-foundation.org>,
	Alan Stern <stern@rowland.harvard.edu>,
	Andrea Parri <parri.andrea@gmail.com>,
	Boqun Feng <boqun.feng@gmail.com>,
	Nicholas Piggin <npiggin@gmail.com>,
	David Howells <dhowells@redhat.com>,
	j alglave <j.alglave@ucl.ac.uk>,
	luc maranget <luc.maranget@inria.fr>, akiyks <akiyks@gmail.com>,
	linux-toolchains <linux-toolchains@vger.kernel.org>,
	linux-arch <linux-arch@vger.kernel.org>
Subject: Re: [RFC PATCH] LKMM: Add ctrl_dep() macro for control dependency
Date: Fri, 1 Oct 2021 12:13:28 -0400 (EDT)	[thread overview]
Message-ID: <2088260319.47978.1633104808220.JavaMail.zimbra@efficios.com> (raw)
In-Reply-To: <20210929174146.GF22689@gate.crashing.org>

----- On Sep 29, 2021, at 1:41 PM, Segher Boessenkool segher@kernel.crashing.org wrote:

> Hi!
> 
> On Wed, Sep 29, 2021 at 02:28:37PM +0200, Florian Weimer wrote:
>> If you need a specific instruction emitted, you need a compiler
>> intrinsic or inline assembly.
> 
> Not an intrinsic.  Builtins (like almost all other code) do not say
> "generate this particular machine code", they say "generate code that
> does <this>".  That is one reason why builtins are more powerful than
> inline assembler (another related reason is that they tell the compiler
> exactly what behaviour is expected).
> 
>> I don't think it's possible to piggy-back this on something else.
> 
> Unless we get a description of what this does in term of language
> semantics (instead of generated machine code), there is no hope, even.

Hi Segher,

Let me try a slightly improved attempt at describing what I am looking
for in terms of language semantics.

First, let's suppose we define two new compiler builtins, e.g.
__sync_ctrl_dep_rw() and __sync_ctrl_dep_acquire().

Their task would be to ensure that a R->W or R->RW (acquire) dependency between the
volatile loads used as input of the evaluated expression and following volatile
stores, volatile loads for R->RW, volatile asm, memory clobbers, is present in the
following situations:

When the builtin is used around evaluation of the left operand of the && (logical
AND) and || (logical OR) expression, the R->W or R->RW dependency should be
present before evaluating the right operand.

When the builtin is used around evaluation of the first operand of the ternary
"question-mark" operator, the R->W or R->RW dependency should be present before
evaluating the second or third operands.

When the builtin is used around evaluation of the controlling expressions of
if, switch, while, and do-while statements, as well as of the second operand of
the for statement, the R->W or R->RW dependency should be present before the
next sequence point is evaluated.

One cheap way to achieve said R->W dependency (as well as R->RW on architectures which
to not reorder R->R) is to ensure that the generated assembly contains a conditional
branch. Other ways to ensure this include more heavy-weight approaches such as explicit
barriers.

Hopefully my description above is slightly closer to the expected language
semantics.

Thanks,

Mathieu

-- 
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com