Hi, Andrew:
> On Jun 30, 2021, at 2:20 PM, Andrew Pinski <pins...@gmail.com> wrote:
>>>>>
>>>>
>>>> Then, where is the correct place for “.DEFERRED_INIT(size2,2)?
>>>>
>>>> The variable “size2” is a block scope variable which is declared inside
>>>> the “if” condition:
>>>
>>> But that's obviously not how it behaves
>>> During into SSA phase since we're inserting a PHI for it - and we're
>>> inserting it because of the use in the DEFERED_INIT call. I suppose you
>>> need to fiddle with the SSA rewrite and avoid treating the use as a use but
>>> only for the purpose of inserting PHIs...
>>
>> Please see my other email on the new small testing case without
>> -ftrivial-auto-var-init. The same issue in SSA with that testing case even
>> without -ftrivial-auto-var-init.
>> It looks like an existing bug to me in SSA.
>>
>> Let me know if I still miss anything
>
> Yes you missed it is unspecified what the value if the auto variable
> is used uninitialized. Isn't that the point of what you are trying to
> fix in the first place?
> So ccp takes PHI<a_1(D), 4> and says since a_1(D) is uninitialized,
> the value is 4.
I don’t think CCP did anything wrong.
I suspect that SSA did something wrong when constructing PHI node for the
block-scope auto variables.
For the following example:
1 extern void bar (int);
2 extern int ART_INIT(int, int);
3 void foo (int a)
4 {
5 int i;
6 for (i = 0; i < a; i++) {
7 if (__extension__({int size2;
8 size2 = ART_INIT (size2, 2);
9 size2 = 4;
10 size2 > 5;}))
11 bar (a);
12 }
13 }
~
Is it legal to propagate constant 4 at line 9 to the first argument of call to
ART_INIT at line 8?
Given size2 is a block-scope auto-variable whose scope is from line 7 to line
10. The value of size2 should not be carried over through
loop iterations from my understanding.
Qing
>
> Thanks,
> Andrew
>
>>
>> Qing
>>>
>>> You might be able to construct a testcase which has a use before the real
>>> init where then the optimistic CCP propagation will defeat the DEFERED_INIT
>>> otherwise.
>>>
>>> I'd need to play with the actual patch to find a good solution to this
>>> problem.
>>>
>>> Richard.
>>>
>>
