[PATCH] D23237: [MSSA] Fix PR28880 by fixing stack popping behavior.

[Daniel Berlin via llvm-commits]

Yes, doing that fixes it on all testcases i have, new patch coming up.


On Sat, Aug 6, 2016 at 10:59 AM, Daniel Berlin <dberlin at dberlin.org> wrote:

> Ugh, it's worse than that.
>
> The push may have happened in an intervening dominator.
>
> Okay, so i can come up with CFG's where this logic is still wrong.
> I think we need to track what memoryAccess the lowerbound was and see if
> the block still dominates us.
> If not, the lower bound is not valid.
>
>
> On Sat, Aug 6, 2016 at 10:24 AM, Daniel Berlin <dberlin at dberlin.org>
> wrote:
>
>> I believe i still have an off by one error here.
>> The lower bounds are in terms of stack places (IE 0... size - 1), and the
>> size is +1 of that.
>> So i think it's missing checking one access sometimes.  I think we need
>> to track size - 1.
>>
>> Checking some testcases.
>>
>>
>>
>> On Sat, Aug 6, 2016 at 9:56 AM, Sebastian Pop <sebpop at gmail.com> wrote:
>>
>>> sebpop added a comment.
>>>
>>> LGTM.
>>> Let's enable gvn-hoist back again once this is committed.
>>>
>>>
>>> ================
>>> Comment at: test/Transforms/Util/MemorySSA/pr28880.ll:54
>>> @@ +53,2 @@
>>> +
>>> +!0 = !{!"Apple LLVM version 8.0.0 (clang-800.0.35)"}
>>> ----------------
>>> You may want to remove all these attributes.
>>>
>>>
>>> https://reviews.llvm.org/D23237
>>>
>>>
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160806/175a2df8/attachment.html>