[PATCH] D108732: [AMDGPU] Invert partial vgpr to agpr spill lane order
Stanislav Mekhanoshin via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Wed Aug 25 15:12:55 PDT 2021
rampitec added a comment.
In D108732#2966072 <https://reviews.llvm.org/D108732#2966072>, @arsenm wrote:
> I don't understand what the problem is. The spilling is all done with 32-bit pieces, so why does the alignment matter?
Say we need to spill v[0:3] like in the added test and we have one agpr available. Current logic will spill v0 into agpr and v[1:3] into scratch.
With mubuf there is no problem, we will issue 3 buffer_store instructions, but with flat scratch it will be a single store of an unaligned register.
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D108732/new/
https://reviews.llvm.org/D108732
More information about the llvm-commits
mailing list