Re: [middle-end PATCH] Prefer PLUS over IOR in RTL expansion of multi-word shifts/rotates.

[Richard Biener <richard.guenther@gmail.com>]

On Fri, Jan 19, 2024 at 2:26 PM Roger Sayle <roger@nextmovesoftware.com> wrote:
>
>
> Hi Richard,
>
> Thanks for the speedy review.  I completely agree this patch
> can wait for stage1, but it's related to some recent work Andrew
> Pinski has been doing in match.pd, so I thought I'd share it.
>
> Hypothetically, recognizing (x<<4)+(x>>60) as a rotation at the
> tree-level might lead to a code quality regression, if RTL
> expansion doesn't know to lower it back to use PLUS on
> those targets with lea but without rotate.
>
> > From: Richard Biener <richard.guenther@gmail.com>
> > Sent: 19 January 2024 11:04
> > On Thu, Jan 18, 2024 at 8:55 PM Roger Sayle <roger@nextmovesoftware.com>
> > wrote:
> > >
> > > This patch tweaks RTL expansion of multi-word shifts and rotates to
> > > use PLUS rather than IOR for disjunctive operations.  During expansion
> > > of these operations, the middle-end creates RTL like (X<<C1) | (Y>>C2)
> > > where the constants C1 and C2 guarantee that bits don't overlap.
> > > Hence the IOR can be performed by any any_or_plus operation, such as
> > > IOR, XOR or PLUS; for word-size operations where carry chains aren't
> > > an issue these should all be equally fast (single-cycle) instructions.
> > > The benefit of this change is that targets with shift-and-add insns,
> > > like x86's lea, can benefit from the LSHIFT-ADD form.
> > >
> > > An example of a backend that benefits is ARC, which is demonstrated by
> > > these two simple functions:
> > >
> > > unsigned long long foo(unsigned long long x) { return x<<2; }
> > >
> > > which with -O2 is currently compiled to:
> > >
> > > foo:    lsr     r2,r0,30
> > >         asl_s   r1,r1,2
> > >         asl_s   r0,r0,2
> > >         j_s.d   [blink]
> > >         or_s    r1,r1,r2
> > >
> > > with this patch becomes:
> > >
> > > foo:    lsr     r2,r0,30
> > >         add2    r1,r2,r1
> > >         j_s.d   [blink]
> > >         asl_s   r0,r0,2
> > >
> > > unsigned long long bar(unsigned long long x) { return (x<<2)|(x>>62);
> > > }
> > >
> > > which with -O2 is currently compiled to 6 insns + return:
> > >
> > > bar:    lsr     r12,r0,30
> > >         asl_s   r3,r1,2
> > >         asl_s   r0,r0,2
> > >         lsr_s   r1,r1,30
> > >         or_s    r0,r0,r1
> > >         j_s.d   [blink]
> > >         or      r1,r12,r3
> > >
> > > with this patch becomes 4 insns + return:
> > >
> > > bar:    lsr     r3,r1,30
> > >         lsr     r2,r0,30
> > >         add2    r1,r2,r1
> > >         j_s.d   [blink]
> > >         add2    r0,r3,r0
> > >
> > >
> > > This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
> > > and make -k check, both with and without --target_board=unix{-m32}
> > > with no new failures.  Ok for mainline?
> >
> > For expand_shift_1 you add
> >
> > +                where C is the bitsize of A.  If N cannot be zero,
> > +                use PLUS instead of IOR.
> >
> > but I don't see a check ensuring this other than mabe CONST_INT_P (op1)
> > suggesting that we enver end up with const0_rtx here.  OTOH why is N zero a
> > problem and why is it not in the optabs.cc case where I don't see any such check
> > (at least not obvious)?
>
> Excellent question.   A common mistake in writing a rotate function in C
> or C++ is to write something like (x>>n)|(x<<(64-n)) or (x<<n)|(x>>(64-n))
> which invokes undefined behavior when n == 0.  It's OK to recognize these
> as rotates (relying on the undefined behavior), but correct/portable code
> (and RTL) needs the correct idiom(x>>n)|(x<<((-n)&63), which never invokes
> undefined behaviour.  One interesting property of this idiom, is that shift
> by zero is then calculated as (x>>0)|(x<<0) which is x|x.  This should then
> reveal the problem, for all non-zero values the IOR can be replaced by PLUS,
> but for zero shifts, X|X isn't the same as X+X or X^X.
>
> This only applies for single word rotations, and not multi-word shifts
> nor multi-word rotates, which explains why this test is only in one place.
>
> In theory, we could use ranger to check whether a rotate by a variable
> amount can ever be by zero bits, but the simplification used here is to
> continue using IOR for variable shifts, and PLUS for fixed/known shift
> values.  The last remaining insight is that we only need to check for
> CONST_INT_P, as rotations/shifts by const0_rtx are handled earlier in
> this function (and eliminated by the tree-optimizers), i.e. rotation by
> a known constant is implicitly a rotation by a known non-zero constant.

Ah, I see.  It wasn't obvious the expmed.cc case was for rotations only.

The patch is OK as-is for stage1 (which also gives others plenty of time
to comment).

I wonder if you can add a testcase though?

Thanks,
Richard.

> This is a little clearer if you read/cite more of the comment that was
> changed.  Fortunately, this case is also well covered by the testsuite.
> I'd be happy to change the code to read:
>
>         (CONST_INT_P (op1) && op1 != const0_rtx)
>         ? add_optab
>         : ior_optab
>
> But the test "if (op1 == const0_rtx)" already appears on line 2570
> of expmed.cc.
>
>
> > Since this doesn't seem to fix a regression it probably has to wait for
> > stage1 to re-open.
> >
> > Thanks,
> > Richard.
> >
> > > 2024-01-18  Roger Sayle  <roger@nextmovesoftware.com>
> > >
> > > gcc/ChangeLog
> > >         * expmed.cc (expand_shift_1): Use add_optab instead of ior_optab
> > >         to generate PLUS instead or IOR when unioning disjoint bitfields.
> > >         * optabs.cc (expand_subword_shift): Likewise.
> > >         (expand_binop): Likewise for double-word rotate.
> > >
>
>
> Thanks again.
>