[llvm] [SLPVectorizer][X86] Free load cost for stores with constant pointers (PR #118016)

[Simon Pilgrim via llvm-commits]

RKSimon wrote:

There's plenty of scope for better vector constant rematerialisation, but gpr->fpu transfer + shuffle perf isn't always great, it avoids memory latency but can cause a lot of throughput + register pressure issues - just trying to do this in DAGISel doesn't work well.

I started the X86FixupVectorConstants pass to begin reducing vector constant pool traffic and there's plenty more to do to remove constant broadcasts etc. in X86ISelLowering, that should then put us in a position where we can do other things as well - similar to how MachineLICM attempts to hoist out constant loads if there's spare registers.

https://github.com/llvm/llvm-project/pull/118016