Re: [PATCH 2/2] tree-optimization/104530 - Mark defs dependent on non-null stale.

[Andrew MacLeod <amacleod@redhat.com>]

On 2/23/22 02:33, Richard Biener wrote:
> On Tue, Feb 22, 2022 at 5:42 PM Andrew MacLeod via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>> This patch simply leverages the existing computation machinery to
>> re-evaluate values dependent on a newly found non-null value
>>
>> Ranger associates a monotonically increasing temporal value with every
>> def as it is defined.  When that value is used, we check if any of the
>> values used in the definition have been updated, making the current
>> cached global value stale.  This makes the evaluation lazy, if there are
>> no more uses, we will never re-evaluate.
>>
>> When an ssa-name is marked non-null it does not change the global value,
>> and thus will not invalidate any global values.  This patch marks any
>> definitions in the block which are dependent on the non-null value as
>> stale.  This will cause them to be re-evaluated when they are next used.
>>
>> Imports: b.0_1  d.3_7
>> Exports: b.0_1  _2  _3  d.3_7  _8
>>            _2 : b.0_1(I)
>>            _3 : b.0_1(I)  _2
>>            _8 : b.0_1(I)  _2  _3  d.3_7(I)
>>
>>      b.0_1 = b;
>>       _2 = b.0_1 == 0B;
>>       _3 = (int) _2;
>>       c = _3;
>>       _5 = *b.0_1;        <<-- from this point b.0_1 is [+1, +INF]
>>       a = _5;
>>       d.3_7 = d;
>>       _8 = _3 % d.3_7;
>>       if (_8 != 0)
>>
>> when _5 is defined, and n.0_1 becomes non-null,  we mark the dependent
>> names that are exports and defined in this block as stale.  so _2, _3
>> and _8.
>>
>> When _8 is being calculated, _3 is stale, and causes it to be
>> recomputed.  it is dependent on _2, alsdo stale, so it is also
>> recomputed, and we end up with
>>
>>     _2 == [0, 0]
>>     _3 == [0 ,0]
>> and _8 = [0, 0]
>> And then we can fold away the condition.
>>
>> The side effect is that _2 and _3 are globally changed to be [0, 0], but
>> this is OK because it is the definition block, so it dominates all other
>> uses of these names, and they should be [0,0] upon exit anyway.  The
>> previous patch ensure that the global values written to
>> SSA_NAME_RANGE_INFO is the correct [0,1] for both _2 and _3.
>>
>> The patch would have been even smaller if I already had a mark_stale
>> method.   I thought there was one, but I guess it never made it in from
>> lack of need at the time.   The only other tweak was to make the value
>> stale if the dependent value was the same as the definitions.
>>
>> This bootstraps on x86_64-pc-linux-gnu with no regressions. Re-running
>> to ensure.
> @@ -1475,6 +1488,15 @@ ranger_cache::update_to_nonnull (basic_block
> bb, tree name)
>          {
>            r.set_nonzero (type);
>            m_on_entry.set_bb_range (name, bb, r);
> +         // Mark consumers of name stale so they can be recomputed.
> +         if (m_gori.is_import_p (name, bb) || m_gori.is_export_p (name, bb))
> +           {
> +             tree x;
> +             FOR_EACH_GORI_EXPORT_NAME (m_gori, bb, x)
> +               if (m_gori.in_chain_p (name, x)
> +                   && gimple_bb (SSA_NAME_DEF_STMT (x)) == bb)
> +                 m_temporal->set_stale (x);
> +           }
>          }
>
> so if we have a BB that exports N names and each of those is updated to nonnull
> this is going to be quadratic?  It also looks like the gimple_bb check
> is cheaper
> than the bitmap test done in in_chain_p.  What comes to my mind is why we need
> to mark "consumers"?  Can't consumers check their uses defs when they look
> at their timestamp?  This whole set_stale thing doesn't seem to be

They do.  The timestamps only look at direct uses. Any use of _2 should 
look at the def and notice it is stale relative to b.0_1 automatically. 
We miss the opportunity in the example which uses _3 to compute _8.  _3 
is directly dependent on _2 whose def is not stale relative to _3, so we 
miss the transitive staleness via b.0_1.   This marks all the consumers 
whose calculation is derived from the now non-null value as stale.   
Within the block, it is fully transitive and anything potentially 
derived from NAME will be recalculated if it is used.  In old EVRP 
terms, it would be like updating the current value vector for any 
ssa-names derived from NAME when it becomes non-null, except it is done 
lazily.


> transitive anyway,
> consider:
>
>     _1 = ...
>
> <bb>
>     _2 = _1 + ..;
>
> <bb>
>    _3 = _2 + ...;
>
> so when _1 is updated to non-null we mark _2 as stale but _3 should
> also be stale, no?
> When we visit _3 before eventually getting to _2 (to see whether it
> updates and thus
> we more precisely we know if it makes _3 stale) we won't re-evaluate it?

> That said, the change looks somewhat ad-hoc to get to 1-level deep second-level
> opportunities?

The patch applies only to dom-walks, and primarily targets definitions 
in the current block that we have already seen that we now know are 
stale. It is one approach to applying non-null later in the same block 
without resorting to much of an algorithmic change.  It's not really 
intended to affect anything cross block as that is handled differently 
via the GORI engine.  It would provide better on-exit ranges in the 
definition block for some of the names involved.

That said, I'm not crazy about putting anything else into this release 
anyway, so if the regressions isn't serious enough, then I'd simply wait 
for the revamp of side-effects in the next release to deal with it.

Andrew