[polly][vectorizer] do not emit bitcast i32** to <1 x i32*>*

[Tobias Grosser]

On 10/17/2013 05:07 AM, Sebastian Pop wrote:
> Tobias Grosser wrote:
>> On 10/15/2013 06:33 PM, Sebastian Pop wrote:
>>> Hi Tobi,
>>>
>>> Sebastian Pop wrote:
>>>> Tobias Grosser wrote:
>>>>> 1) Bitcast patch
>>>>>
>>>>> The patch goes definitely in the right direction. However, it would
>>>>> be nice if you could besides modifying the existing test cases also
>>>>> provide a test cases that fail without these changes.
>>>>
>>>> I will reduce a testcase and repost the patch for review.
>>>
>>> Here is a C reduced testcase: (thanks to http://embed.cs.utah.edu/creduce/ )
>>
>> Thanks.
>>
>>> I was not able to write a Polly specific testcase that would fail with opt.
>>> Tobi, can you help extracting a testcase?
>>
>> Yes, attached.
>>
>> Run it with:
>>
>> polly-opt -O3 -polly -polly-vectorizer=polly -polly-ignore-aliasing
>>
>
> See attached the amended patch with this testcase.  Ok to commit?

Sorry, no.

Having the '-O3' in the test case makes it fragile to future changes, 
unnecessary slow and also difficult to understand the actual problem 
that is tested. We should reduce this to the list of passes.

Secondly, as written before, "It would be nice if we can change [the 
test case] such that the '-polly-ignore-aliasing' flag would not even be 
needed."

I did not fully reduce the test case yet, but only tried to give you an 
LLVM-IR level one quickly to unblock your work. Let me know if you need 
help reducing the list of passes or changing the code to do not require 
-polly-ignore-aliasing.

Cheers,
Tobias