Re: [PATCH] Loop unswitching: support gswitch statements.

["Martin Liška" <mliska@suse.cz>]

On 12/2/21 13:01, Richard Biener wrote:
> On Thu, Dec 2, 2021 at 12:45 PM Martin Liška <mliska@suse.cz> wrote:
>>
>> On 12/1/21 19:21, Andrew MacLeod wrote:
>>> On 12/1/21 09:48, Martin Liška wrote:
>>>> On 12/1/21 15:34, Richard Biener wrote:
>>>>> On Wed, Dec 1, 2021 at 3:25 PM Martin Liška <mliska@suse.cz> wrote:
>>>>>>
>>>>>> On 12/1/21 15:19, Richard Biener wrote:
>>>>>>> which is compute the range of 'lhs' on edge_true into predicate->true_range,
>>>>>>> assign that same range to ->false_range and then invert it to get the
>>>>>>> range on the false_edge.  What I am saying is that for better precision
>>>>>>> you should do
>>>>>>>
>>>>>>>         ranger->range_on_edge (predicate->false_range, edge_false, lhs);
>>>>>>>
>>>>>>> rather than prematurely optimize this to the inversion of the true range
>>>>>>> since yes, ranger is CFG sensitive and only the_last_ predicate on a
>>>>>>> long CFG path is actually inverted.
>>>>>>>
>>>>>>> What am I missing?
>>>>>>
>>>>>> I might be misunderstood, but I think it's the problem defined here:
>>>>>> https://gcc.gnu.org/pipermail/gcc-patches/2021-November/584605.html
>>>>>>
>>>>>> where I used the ranger->range_on_edge on the false_edge.
>>>>>
>>>>> Ah, OK.  But then even the true_edge range is possibly wrong, no?
>>>>
>>>> You are of course correct, I've just proved that in debugger ://
>>>>
>>>>> Consider
>>>>>
>>>>>     for (;;)
>>>>>        {
>>>>>            if (a < 100)
>>>>>              if (a > 50)  // unswitch on this
>>>>>                /* .. */
>>>>>            if (a < 120)
>>>>>                /* ... */
>>>>>        }
>>>>>
>>>>> then you record [51, 99] for true_range of the a > 50 predicate and thus
>>>>> simplification will simplify the if (a < 120) check, no?
>>>>
>>>> Yep.
>>>>
>>>>>
>>>>> You can only record the range from the (CFG independent) a > 50 check,
>>>>> thus [51, +INF] but of course at simplification time you can also use
>>>>> the CFG context at each simplification location.
>>>>
>>>> @Andrew: How can I easily get irange based just on a stmt? Something like fold_range
>>>> with int_range_max as the 3rd argument?
>>>>
>>> Sorry, I miss these things if I'm not directly CC'd a lot :-)
>>>
>>> So you just want to know the basic range the stmt generates without context?    Sure, what you say would be fine, but your want to initialize it to the type range:
>>
>> Yes, I want to know range of LHS in a gcond statement and the same for cases in a gswitch statement.
>>
>>>
>>> int_range_max range (TREE_TYPE (name));
>>>
>>> you can also simply trigger it using the current SSA_NAME_RANGE_INFO global  values query instead of the default current contextual one... which , if there isnt a global range, will automatically use the range of the type of the argument.. so maybe just try
>>>
>>> fold_range (r, stmt, get_global_range_query ())
>>
>> Doing
>>
>>         predicate->true_range = int_range_max (TREE_TYPE (lhs));
>>         fold_range (predicate->true_range, stmt, get_global_range_query ());
>>         predicate->true_range.debug();
>>
>> gives me _Bool VARYING.
> 
> Likely because that gives a range for the bool result rather than
> a range for the LHS of a LHS op RHS on the true or false edge.

Yes :) I guess Andrew can help us.

> I would guess no stmt level API gives that.  In previous VRP
> incarnation assert expr discovery would yield the asserts
> that hold for the LHS on the respective edge and from the asserts
> ranges could be determined.

About the gswitch statements, I need similarly irange for a switch statement
on an edge e.

Thanks for help,
Martin

> 
> Richard.
> 
>>
>> Martin
>>
>>>
>>> Andrew
>>>
>>>
>>>
>>>
>>