[PATCH 2/3] ARM cost model: Address computation in vector mem ops not free
Arnold
aschwaighofer at apple.com
Thu Feb 7 05:20:46 PST 2013
I has a compound effect on throughput if you can only issue three a cycle. So there will be an impact on four vs one of them. This is what I am trying to capture at a high Level.
Sent from my iPhone
On Feb 6, 2013, at 11:55 PM, Nadav Rotem <nrotem at apple.com> wrote:
> I am not sure that its worth modeling this because it only affects the latency and not the throughput of the machine.
>
> On Feb 5, 2013, at 3:33 PM, Arnold Schwaighofer <aschwaighofer at apple.com> wrote:
>
>> Updated patch. We now add the cost of address computation as part of the memory instruction cost.
>>
>>
>> Thanks
>>
>> <0001-ARM-cost-model-Address-computation-in-vector-mem-ops.patch>
>>
>> On Feb 1, 2013, at 1:39 PM, Renato Golin <renato.golin at linaro.org> wrote:
>>
>>> On 1 February 2013 18:07, Nadav Rotem <nrotem at apple.com> wrote:
>>>> The problem is that we decide on the kind of GEP to use only when we vectorize the load/stores. This happens in vectorizeMemoryInstruction. I think that we need to fix this in LoopVectorizationCostModel::getInstructionCost in the load/store switch cases. We have code for checking if the load/store is wide or if it is scalarized.
>>>
>>> Good point! Shouldn't be too hard, though.
>>>
>>> cheers,
>>> --renato
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20130207/36a29a23/attachment.html>
More information about the llvm-commits
mailing list