Re: [PATCH] Fixup possible VR_ANTI_RANGE value_range issues

[Aldy Hernandez <aldyh@redhat.com>]

On 2/28/23 10:41, Richard Biener wrote:
> On Mon, 27 Feb 2023, Aldy Hernandez wrote:
> 
>>
>>
>> On 2/27/23 14:58, Richard Biener wrote:
>>> After fixing PR107561 the following avoids looking at VR_ANTI_RANGE
>>> ranges where it doesn't seem obvious the code does the correct
>>> thing here (lower_bound and upper_bound do not work as expected).
>>
>> I do realize there's some confusion here, and some of it is my fault. This has
>> become obvious in my upcoming work removing all of legacy.
>>
>> What's going on is that ultimately min/max are broken with (non-legacy)
>> iranges.  Or at the very least inconsistent between legacy and non-legacy.
>> These are left over from the legacy world, and have been marked DEPRECATED for
>> a few releases, but the middle end warnings continued to use them and even
>> added new uses after they were obsoleted.
>>
>> min/max have different meanings depending on kind(), which is also deprecated,
>> btw.  They are the underlying min/max fields from the legacy implementation,
>> and thus somewhat leak the implementation details. Unfortunately, they are
>> being called from non-legacy code which is ignoring the kind() field.
>>
>> In retrospect I should've converted everything away from min/max/kind years
>> ago, or at the very least converted min/max to work with non-legacy more
>> consistently.
>>
>> For the record:
>>
>>    enum value_range_kind kind () const;		// DEPRECATED
>>    tree min () const;				// DEPRECATED
>>    tree max () const;				// DEPRECATED
>>    bool symbolic_p () const;			// DEPRECATED
>>    bool constant_p () const;			// DEPRECATED
>>    void normalize_symbolics ();			// DEPRECATED
>>    void normalize_addresses ();			// DEPRECATED
>>    bool may_contain_p (tree) const;		// DEPRECATED
>>    bool legacy_verbose_union_ (const class irange *);	// DEPRECATED
>>    bool legacy_verbose_intersect (const irange *);	// DEPRECATED
>>
>> In my local branch I tried converting all the middle-end legacy API uses to
>> the new API, but a bunch of tests started failing, and lots more false
>> positives started showing up in correct code.  I suspect that's part of the
>> reason legacy continued to be used in these passes :-/.  As you point out in
>> the PR, the tests seem designed to test the current (at times broken)
>> implementation.
>>
>> That being said, the 5 fixes in your patch are not wrong to begin with,
>> because all uses guard lower_bound() and upper_bound() which work correctly.
>> They return the lowest and uppermost possible bound for the range (ignoring
>> the underlying implementation).  So, the lower bound of a signed non-zero is
>> -INF because ~[0,0] is really [-INF,-1][1,+INF]. In the min/max legacy code,
>> min() of signed non-zero (~[0,0]) is 0.  The new implementation has no concept
>> of anti-ranges, and we don't leak any of that out.
>>
>> Any uses of min/max without looking at kind() are definitely broken. OTOH uses
>> of lower/upper_bound are fine and should work with both legacy and non-legacy.
>>
>> Unrelated, but one place where I haven't been able to convince myself that the
>> use is correct is bounds_of_var_in_loop:
>>
>>> /* Check if init + nit * step overflows.  Though we checked
>>>         scev {init, step}_loop doesn't wrap, it is not enough
>>>         because the loop may exit immediately.  Overflow could
>>>         happen in the plus expression in this case.  */
>>>      if ((dir == EV_DIR_DECREASES
>>>           && compare_values (maxvr.min (), initvr.min ()) != -1)
>>>          || (dir == EV_DIR_GROWS
>>>       && compare_values (maxvr.max (), initvr.max ()) != 1))
>>
>> Ughh...this is all slated to go away, and I have patches removing all of
>> legacy and the old API.
>>
>> Does this help?  Do you still think lower and upper bound are not working as
>> expected?
> 
> lower_bound/upper_bound work as expected,
> tree_lower_bound/tree_upper_bound do not.  I've checked and all uses
> I "fixed" use lower_bound/upper_boud.
> 
> tree_lower_bound & friends are 'protected', but in the above light the
> comment
> 
>    // potential promotion to public?
> 
> looks dangerous (I was considering using them ...).

First of all, my apologies for the time you've spent on this.  This 
needed cleaning up a few releases ago, but with legacy in place, it kept 
getting pushed further away.  This is my first order of business once 
stage1 opens.  That being said, thank you for spot checking this.

It looks like tree_lower_bound was just syntactic sugar for m_base[pair 
* 2].  The comment is likely wrong.  It should've stayed protected, or 
nuked in favor of accessing m_base directly to avoid confusion.

> 
> I'm dropping the patch, it's probably time to work on getting rid of
> 'value_range' uses (aka legacy_mode_p ()) and/or replace raw accesses
> to the min/max for that mode with m_base accesses?  At least
> that there's tree_{upper,lower}_bound for pair != 0 suggests this
> function is used with differing semantics internally which is a

My local tree has a few dozen patches removing legacy and converting the 
trees to wide_int.  In my work, legacy_mode_p, min, max, kind, 
tree_lower_bound, etc are all gone.  I can make the branch public if you 
like.

> source of confusion (to me at least).  tree_{lower,upper}_bound
> should be able to assert it's not a legacy mode range and
> min/max should be able to assert that it is.

Agreed.  One possible problem *may be* vr-values.cc which still uses 
min() and max() throughout can be called on legacy as well as non-legacy.

This unfortunately means that since kind/min/max will return different 
results for legacy/non-legacy (kind() never returns VR_ANTI_RANGE for 
non-legacy), simplify_using_ranges can theoretically give different 
results depending on how it's called.  In practice, only DOM/VRP call 
simplify_using_ranges and always with non-legacy.  However, 
simplify_using_ranges uses get_value_range() internally which uses 
legacy.  Sorry, I just realized this now while auditing the code.

I honestly have no idea if this matters in practice, and since this has 
been this way for a few releases, I'm afraid of touching it so late in 
the cycle.  But I'm happy to help if you think it must be fixed in this 
release.

Hmmmm, it looks like there are just a few non-legacy uses in 
vr-values.cc (int_range<2> and int_range_max), and never in code paths 
checking min/max/kind.  So your suggestions about the asserts may be all 
we need, dunno.

> 
> Btw, legacy_upper_bound looks wrong:
> 
>    if (m_kind == VR_ANTI_RANGE)
>      {
>        tree typ = type (), t;
>        if (pair == 1 || vrp_val_is_min (min ()))
>          t = vrp_val_max (typ);
>        else
>          t = wide_int_to_tree (typ, wi::to_wide (min ()) - 1);
> 
> not sure what the pair == 1 case is about, but the upper bound
> of an anti-range is only special when max() is vrp_val_is_max,
> aka ~[low, +INF] where it then is low - 1, but the above checks
> for ~[-INF, high] and returns +INF in that case.  That's correct
> unless high == +INF at least, but the else case is wrong for
> all high != +INF?

An anti-range is really two pairs.  So, ~[6,9] is really:

[-MIN, 5][10, MAX]

Pair 0 is [-MIN, 5] and pair 1 is [10, MAX].  This means that the upper 
bound for the pair==1 case is just MAX.

> 
> I think it should be
> 
>        if (pair == 1 || !vrp_val_is_max (max ()))
>          t = vrp_val_max (typ);
>        else
>          t = wide_int_to_tree (typ, wi::to_wide (min ()) - 1);
> 
> no?

I think both of us are wrong.  The original vrp_val_is_min(min()) call 
will always return false.  And your !vrp_val_is_max(max()) is always 
true.  This is because irange::set() will normalize anti-ranges into a 
VR_RANGE when possible, in order to avoid multiple representations for 
the same range:

       else if (is_min)
         {
	  tree one = build_int_cst (TREE_TYPE (max), 1);
	  min = int_const_binop (PLUS_EXPR, max, one);
	  max = vrp_val_max (TREE_TYPE (max));
	  kind = VR_RANGE;
         }
       else if (is_max)
         {
	  tree one = build_int_cst (TREE_TYPE (min), 1);
	  max = int_const_binop (MINUS_EXPR, min, one);
	  min = vrp_val_min (TREE_TYPE (min));
	  kind = VR_RANGE;
         }

So ~[-MIN,5] is represented as [6,MAX] and ~[5,+MAX] as [-MIN,4].

Ughhh... so it looks like legacy_upper_bound is correct, but the 
vrp_val_is_min() check is useless.

Am I missing something?
Aldy