[AMDGPU] Revert wide LDS DMA support.
authorStanislav Mekhanoshin <Stanislav.Mekhanoshin@amd.com>
Wed, 11 May 2022 20:04:17 +0000 (13:04 -0700)
committerStanislav Mekhanoshin <Stanislav.Mekhanoshin@amd.com>
Mon, 16 May 2022 18:23:35 +0000 (11:23 -0700)
commit332b73fe12c6b383168e03eddd92020d38bc7583
tree35677589e536a7ece4b031306338dfd6b166c264
parent5edd7665fd163555f1afe2187907eba95d24c11b
[AMDGPU] Revert wide LDS DMA support.

This reverts ffbee7acdcaaf, see also bug 37653 which it was fixing.
The bug claims this is an undocumented feature which actually works.
In the reality it is documented as not working for a good reason.
It likely does something, but it is useless anyway. These instructions
write into the LDS. The LDS address is:

M0 + inst_offset + (TIDinWave * 4).

For a store wider than a DWORD neighboring lanes will overwrite each
other.

Differential Revision: https://reviews.llvm.org/D125409
llvm/lib/Target/AMDGPU/BUFInstructions.td
llvm/test/MC/AMDGPU/mubuf.s