[PATCH] D123634: [AMDGPU] Split unaligned 4 DWORD DS operations

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/SIInstrInfo.td:591-595
+def load_align_less_than_4_local : PatFrag<(ops node:$ptr),
+                                           (load_local node:$ptr)> {
+  let IsLoad = 1;
+  let IsNonExtLoad = 1;
+}
----------------
arsenm wrote:
>  Do you really need these PatFrags? The legalizer made the split decisions and the patterns shouldn't need logic deper than it's able to select (which it can regardless of the alignment)
If I am using just load_local/store_local I need to override selection of the ds_read2_b64 for alignments less than 4. Using AddedComplexity does not help, it shall be selected differently for alignment less than 4 and 4.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D123634/new/

https://reviews.llvm.org/D123634