[X86] AMD Jaguar scheduler doesn't correctly model 256-bit AVX instructions

[RKSimon]

Bugzilla Link	28573
Resolution	FIXED
Resolved on	Jan 19, 2018 04:31
Version	trunk
OS	Windows NT
Blocks	#31672
CC	@filcab,@RKSimon,@silvasean,@rotateright

Extended Description

All current AMD CPUs capable of AVX 'double pump' 256-bit vector operations through 128-bit vector ALUs.

On Jaguar this means while the latency is typically (but not always) the same for a xmm and ymm variants, the throughput is half what is predicted by the scheduler/cost-models. This needs to be accounted for.

[RKSimon]

https://reviews.llvm.org/D33099

[RKSimon]

Resolving this, all btver2 AVX ymm instructions now show the 'double pumping' half throughput of their xmm equivalents.

[RKSimon]

mentioned in issue #31672