Re: [dpdk-dev] [PATCH] i40e: improve performance of vector PMD

[Bruce Richardson <bruce.richardson@intel.com>]

On Thu, Apr 14, 2016 at 11:15:21AM +0100, Bruce Richardson wrote:
> An analysis of the i40e code using Intel® VTune™ Amplifier 2016 showed
> that the code was unexpectedly causing stalls due to "Loads blocked by
> Store Forwards". This can occur when a load from memory has to wait
> due to the prior store being to the same address, but being of a smaller
> size i.e. the stored value cannot be directly returned to the loader.
> [See ref: https://software.intel.com/en-us/node/544454]
> 
> These stalls are due to the way in which the data_len values are handled
> in the driver. The lengths are extracted using vector operations, but those
> 16-bit lengths are then assigned using scalar operations i.e. 16-bit
> stores.
> 
> These regular 16-bit stores actually have two effects in the code:
> * they cause the "Loads blocked by Store Forwards" issues reported
> * they also cause the previous loads in the RX function to actually be a
> load followed by a store to an address on the stack, because the 16-bit
> assignment can't be done to an xmm register.
> 
> By converting the 16-bit stores operations into a sequence of SSE blend
> operations, we can ensure that the descriptor loads only occur once, and
> avoid both the additional store and loads from the stack, as well as the
> stalls due to the second loads being blocked.
> 
> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> 
Self-NAK on this version. The blend instruction used is SSE4.1 so breaks the
"default" build.

Two obvious options to fix this:
1. Keep the old code with SSE4.1 #ifdefs separating old and new
2. Update the vpmd requirement to SSE4.1, and factor that in during runtime
select of the RX code path.

Personally, I prefer the second option. Any objections?

/Bruce