[patch] fix loopvectorizer bug with multiple address spaces PR16306

[Pekka Jääskeläinen]

OK,

(Sorry for the wrong mailing list.)

Committed the fix. It seems to me that the only place where the address space
can get lost is when bitcasting the original scalar ptr to a vector one. Is this
right?

What shall be the protocol on getting fixes to the maintenance release?
Should I just commit the fix to the 3.3 branch as well (after LLVM 3.3
is actually out)?

On 06/17/2013 09:01 PM, Nadav Rotem wrote:
> Pekka,
>
> LGTM.  Please also make sure that we are not dropping the address space for Loads.
>
> Thanks,
> Nadav
>
> On Jun 17, 2013, at 10:57 AM, Eli Friedman <eli.friedman at gmail.com
> <mailto:eli.friedman at gmail.com>> wrote:
>
>> Forwarding to the correct list.
>>
>> ---------- Forwarded message ----------
>> From:*Pekka Jääskeläinen*<pekka.jaaskelainen at tut.fi
>> <mailto:pekka.jaaskelainen at tut.fi>>
>> Date: Mon, Jun 17, 2013 at 8:06 AM
>> Subject: [patch] fix loopvectorizer bug with multiple address spaces PR16306
>> To:cfe-commits at cs.uiuc.edu <mailto:cfe-commits at cs.uiuc.edu>
>>
>>
>> Hello,
>>
>> Attached patch fixes a severe regression caused by LoopVectorizer when
>> vectorizing loops with memory accesses to non-zero address spaces. It
>> simply drops the AS info. It's a rather nasty problem as there is no
>> way (at least known to me) to disable the loop vectorizer *only* from
>> clang/opt command line.
>>
>> Seems like an obvious fix.
>>
>> http://llvm.org/bugs/show_bug.__cgi?id=16306
>> <http://llvm.org/bugs/show_bug.cgi?id=16306>
>>
>> The patch applies to trunk and 3.3 branches. Tom, could you apply
>> this to the (upcoming?) 3.3 maintenance branch as well?
>>
>> Thanks,
>> --
>> --Pekka
>>
>> _______________________________________________
>> cfe-commits mailing list
>> cfe-commits at cs.uiuc.edu <mailto:cfe-commits at cs.uiuc.edu>
>> http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
>>
>>
>> <loopvec-address-space-fix.patch>
>


-- 
--PJ