[PATCH] D113778: [AMDGPU] Use shift for b64 mov

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec added a comment.

In D113778#3127677 <https://reviews.llvm.org/D113778#3127677>, @sebastian-ne wrote:

> In D113778#3127676 <https://reviews.llvm.org/D113778#3127676>, @arsenm wrote:
>
>> In D113778#3127670 <https://reviews.llvm.org/D113778#3127670>, @foad wrote:
>>
>>> In D113778#3127658 <https://reviews.llvm.org/D113778#3127658>, @arsenm wrote:
>>>
>>>> 64-bit shifts were quarter rate instructions last I checked, so this is slower
>>>
>>> The Write64Bit definitions in SISchedule.td suggest they are half rate on most subtargets and full rate on gfx90a.
>>
>> I think that's probably wrong. Comments in performShlCombine for example say it's quarter rate
>
> It seems to be quarter rate (or something slow) on gfx9, full rate on gfx90a and half rate on gfx10?
> Then it would be worth using on gfx90a and gfx10+.

You do not need this on gfx90a because there is pk_mov. It is arguably the same performance as 2 moves on gfx10.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D113778/new/

https://reviews.llvm.org/D113778