[cfe-commits] [PATCH] Optimize vec3 loads/stores

[Chris Lattner]

On Jul 27, 2012, at 11:08 AM, Hal Finkel wrote:

>> load/stores makes sense in that context. I can only enable this if
>> its OpenCL, but I think anyone who uses vec3 would want this
>> performance win.
> 
> Interesting thanks! I was just wondering why there is a performance win
> at all. I assume that vectors of length 3 are promoted to length-4
> vectors during legalization currently. Why should declaring them to be
> length 4 instead of length 3 affect the code quality at all?

Legalize is basic block local, causing it to miss many important optimizations, just because they happen to span blocks.

-Chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20120727/5ae4a403/attachment.html>