Re: [PATCH 2/2] VR-VALUES: Rewrite test_for_singularity using range_op_handler

[Jeff Law <jeffreyalaw@gmail.com>]

On 9/5/23 01:12, Andrew Pinski wrote:
> On Mon, Sep 4, 2023 at 11:06 PM Jeff Law via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>>
>>
>>
>> On 9/1/23 11:30, Andrew Pinski via Gcc-patches wrote:
>>> So it turns out there was a simplier way of starting to
>>> improve VRP to start to fix PR 110131, PR 108360, and PR 108397.
>>> That was rewrite test_for_singularity to use range_op_handler
>>> and Value_Range.
>>>
>>> This patch implements that and
>>>
>>> OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.
>>>
>>> gcc/ChangeLog:
>>>
>>>        * vr-values.cc (test_for_singularity): Add edge argument
>>>        and rewrite using range_op_handler.
>>>        (simplify_compare_using_range_pairs): Use Value_Range
>>>        instead of value_range and update test_for_singularity call.
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>>        * gcc.dg/tree-ssa/vrp124.c: New test.
>>>        * gcc.dg/tree-ssa/vrp125.c: New test.
>>> ---
>>
>>> diff --git a/gcc/vr-values.cc b/gcc/vr-values.cc
>>> index 52ab4fe6109..2474e57ee90 100644
>>> --- a/gcc/vr-values.cc
>>> +++ b/gcc/vr-values.cc
>>> @@ -904,69 +904,33 @@ simplify_using_ranges::simplify_bit_ops_using_ranges
>>>    }
>>>
>>>    /* We are comparing trees OP1 and OP2 using COND_CODE.  OP1 has
>>> -   a known value range VR.
>>> +   a known value range OP1_RANGE.
>>>
>>>       If there is one and only one value which will satisfy the
>>> -   conditional, then return that value.  Else return NULL.
>>> -
>>> -   If signed overflow must be undefined for the value to satisfy
>>> -   the conditional, then set *STRICT_OVERFLOW_P to true.  */
>>> +   conditional on the EDGE, then return that value.
>>> +   Else return NULL.  */
>>>
>>>    static tree
>>>    test_for_singularity (enum tree_code cond_code, tree op1,
>>> -                   tree op2, const value_range *vr)
>>> +                   tree op2, const int_range_max &op1_range, bool edge)
>>>    {
>>> -  tree min = NULL;
>>> -  tree max = NULL;
>>> -
>>> -  /* Extract minimum/maximum values which satisfy the conditional as it was
>>> -     written.  */
>>> -  if (cond_code == LE_EXPR || cond_code == LT_EXPR)
>>> +  /* This is already a singularity.  */
>>> +  if (cond_code == NE_EXPR || cond_code == EQ_EXPR)
>>> +    return NULL;
>> I don't think this is necessarily the right thing to do for NE.
>>
>> Consider if op1 has the range [0,1] and op2 has the value 1.  If the
>> code is NE, then we should be able to return a singularity of 0 since
>> that's the only value for x where x ne 1 is true given the range for x.
> 
> The "false" edge singularity is already known when NE is supplied. I
> don't think changing it to the "true" edge singularity will be helpful
> all of the time; preferring the value of 0 is a different story.
> But that is a different patch and for a different location rather than
> inside VRP; it should be in either isel or expand (more likely isel).
I forgot something critically important here.  Specifically, this code 
is supposed to be subsumed by Ranger.

The whole point of this routine is to rewrite to EQ/NE comparisons so 
that we can expose equivalences on the true/false arm of the conditional 
(and NE is just as important as EQ).  It's not really about preferring 
any particular value like 0.

The problem with this routine is it loses information after the code has 
been transformed.  Let's say we had a test x < 4.  If we assume that VRP 
is able to prove X doesn't have any of the values [MIN..3], then we can 
change the test to x == 4 and propagate 4 for any uses of X in the true arm.

But on the false ARM we end up with x != 4 which is a wider range than 
[5..MAX].  So if we were to instantiate a new Ranger after the 
transformation we'd lose information on the false arm.  More 
importantly, I think the transformation was bad for either SCEV or loop 
iteration analysis.

When Andrew, Aldy and I kicked this problem around the consensus was 
that Ranger should find and propagate the equivalence, including making 
it visible to jump threading.  That should make the rewriting totally 
unnecessary.

So the net is we really ought to be doing here is looking for cases 
where this code actually helps code generation and if it does we need to 
understand how/why as this code is supposed to go away.

Given you're already poking around in here, you might have such cases 
handy :-)  If you do, I'm sure Andrew, Aldy and I would love to see them.

jeff