Re: [PATCH] PR tree-optimization/31178 - Add rshift side effect.

[Andrew MacLeod <amacleod@redhat.com>]

On 5/18/22 02:41, Richard Biener wrote:
> On Tue, May 17, 2022 at 8:41 PM Andrew MacLeod via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>> This patch implements side effects of the second operand of a shift
>> operation.
>>
>> given A >> B or A << B, the range of B is restricted to [0, PRECISION_A).
>>
>> Fortran is currently more permissive than this, allowing the range to be
>> [0, PRECISION_A], so this si the value we currently default to in this
>> patch.  If the fortran front end were adjusted, we could adjust the end
>> point.
>>
>> This currently bootstraps with no regressions on x86_64-pc-linux-gnu.
>>
>> Is this sufficient, or should I also be checking some other flags which
>> may allow other values outside this range to be valid?
> I think the "undefined behavior" side-effects are dangerous since
> I know of no code that makes sure to clamp the shift argument when
> hoisting it.  sanitizing shifts will also not help to discover such latent
> issues since sanitizing is happening early and it will most definitely
> avoid the hoisting itself.
>
> As to that we _definitely_ want a way to disable this [assumption
> that the shift operand is in-range] if we make that assumption
> even on IL state after optimizations.
>
> Candidates to look for are invariant motion, ifcombine,
> partial PRE, PRE in general (we possibly hoist such expressions
> across function calls that might terminate the program normally).
>
> That is, what prevents
>
>     if (n > 32)
>       abort ();
>     x = i << n;
>
> to be transformed to
>
>     x = i << n;
>     if (n > 32)
>       abort ();
>
> ?  Yes, that's probably a latent issue in some sense but you would
> now optimize the if (n > 32) abort () away while previously x would
> have an undetermined value but we'd still abort.
>
> Do you have some statistics on how this particular side-effect
> improves code generation?
>
> Richard.

None whatsoever...  just a PR requesting it that has been open for 15 
years :-)

If we don't want to do this optimization, then we should kill the PR. I 
can also attach the path to the PR than if anyone cares enough about 
this functionality, they could pursue the other effects...

Andrew