Re: [PATCH] RFC: problems with cast optimization at evaluation time

From: Linus Torvalds <torvalds@linux-foundation.org>
To: Luc Van Oostenryck <luc.vanoostenryck@gmail.com>
Cc: Sparse Mailing-list <linux-sparse@vger.kernel.org>
Subject: Re: [PATCH] RFC: problems with cast optimization at evaluation time
Date: Sat, 2 Jan 2021 11:52:37 -0800	[thread overview]
Message-ID: <CAHk-=wiYyR8gLPPBTwbDf97GiJLOH+EEQED7ARcXR2xyhK4r-w@mail.gmail.com> (raw)
In-Reply-To: <20210102100934.28459-1-luc.vanoostenryck@gmail.com>

On Sat, Jan 2, 2021 at 2:09 AM Luc Van Oostenryck
<luc.vanoostenryck@gmail.com> wrote:
>
> It's easy enough to add a test to only allow this on 'true' integer types
> like done in the patch.

We already linearize casts to different ops for the different cases
(integer/pointer/fp/bool). So yeah, we should do it at that level.

> Again it's easy enough to add some checks but I wonder if all this
> is really worth.

I agree. Remove it from the evaluation phase, and add it to the
linearized SSA format instead.

Note that even on RISC architectures that don't even _have_ truncated
operations (ie an "add" or "not" is always done on the whole
register), doing the truncation to minimal bits is worth it, because
it ends up being just a "upper bits don't matter" and will remove a
_lot_ of sign-extension or zero-extension crud.

So I think the simplification should be done, but I agree it's done in
the wrong place.

Once you do it at the instruction simplification level, all of the
confusion goes away, because a "cast to floating point" is a
fundamentally different operation than "zero-extend integer" at that
level.

            Linus