Add scheduler classes for vector masked load/store operations

[RKSimon]

Bugzilla Link	32857
Resolution	FIXED
Resolved on	May 08, 2018 05:39
Version	trunk
OS	All
Blocks	#31672 #36479 #36256
CC	@filcab,@rotateright
Fixed by commit(s)	rL331760

Extended Description

The vector mask load/store operations (VMASKMOVPD/VMASKMOVPS etc.) often have very different latency/throughput characteristics to basic load/store instructions.

[llvmbot]

It seems we already have these instrs in trunk. I think we could close this record.

https://reviews.llvm.org/rL317196

[RKSimon]

It seems we already have these instrs in trunk. I think we could close this
record.

https://reviews.llvm.org/rL317196

Those are regexp overloads on specific models, what I was suggesting was we improve the current WriteLoad/WriteStore SchedWrite to others such as:

WriteMaskedVectorLoad
WriteMaskedVectorStore

and maybe at the same time:

WriteMaskedVectorGather
WriteMaskedVectorScatter

[RKSimon]

Support for masked loads/stores was added at rL331760.

[RKSimon]

mentioned in issue #36256

[RKSimon]

mentioned in issue #36479