[llvm] r190916 - Lift alignment restrictions for load/store folding on VINSERTF128/VEXTRACTF128. Fixes PR17268. [PATCH]

[Arnold Schwaighofer]

I am okay with this (http://llvm.org/viewvc/llvm-project?view=revision&revision=196294) going into 3.4.

Bill?


On Dec 3, 2013, at 2:30 PM, Pekka Jääskeläinen <pekka.jaaskelainen at tut.fi> wrote:

> Thanks!
> 
> Should I commit this to LLVM 3.4?
> 
> On 12/03/2013 06:42 PM, Arnold Schwaighofer wrote:
>> Committed in r196294.
>> 
>> I chose to put the logic of disabling the vectorizer into opt, like Hal initially suggested. I think it makes more sense to push the flags higher up into the tools. The pass manager builder should be configured by setting its fields (Builder.LoopVectorize, etc.) . Otherwise, we end up with horrible logic in the pass manager builder.
>> 
>> 
>> On Dec 3, 2013, at 7:57 AM, Pekka Jääskeläinen <pekka.jaaskelainen at tut.fi> wrote:
>> 
>>> Ping.
>>> Is there something I can do to get this moving forward to LLVM 3.4?
>>> 
>>> On 11/22/2013 03:57 PM, Pekka Jääskeläinen wrote:
>>>> OK,
>>>> 
>>>> So this will enable to control the vectorizers enabled in opt -O3 by
>>>> adding -fno-slp-vectorizer or similar? Sounds good to me. I propose
>>>> including it also in 3.4 as it causes major headaches in our
>>>> TCE target.
>>>> 
>>>> On 11/21/2013 08:29 PM, Arnold Schwaighofer wrote:
>>>>> 
>>>>> On Nov 21, 2013, at 12:18 PM, Hal Finkel <hfinkel at anl.gov> wrote:
>>>>> 
>>>>>> ----- Original Message -----
>>>>>>> From: "Arnold Schwaighofer" <aschwaighofer at apple.com>
>>>>>>> To: "Hal Finkel" <hfinkel at anl.gov>
>>>>>>> Cc: "Commit Messages and Patches for LLVM"
>>>>>>> <llvm-commits at cs.uiuc.edu>, "Nadav Rotem" <nrotem at apple.com>
>>>>>>> Sent: Thursday, November 21, 2013 11:54:05 AM
>>>>>>> Subject: Re: [llvm] r190916 - Lift alignment restrictions for
>>>>>>> load/store folding on VINSERTF128/VEXTRACTF128. Fixes
>>>>>>> PR17268. [PATCH]
>>>>>>> 
>>>>>>> Oh, I see. We used to use the cl::opt flag for communicating settings
>>>>>>> from the front-end.
>>>>>>> 
>>>>>>> Now, we want it to override settings from the front end (be it opt or
>>>>>>> clang):
>>>>>>> 
>>>>>>> So we want something like:
>>>>>> 
>>>>>> Yes. Although we may want to check for getNumOccurrences() > 0
>>>>>> instead of getPosition() just to be clearer.
>>>>> 
>>>>> Yes. Absolutely.
>>>> 
>>> 
>>> 
>>> --
>>> --Pekka
>> 
> 
> 
> -- 
> --Pekka