[llvm-branch-commits] [llvm] [AMDGPU] Price under-aligned memory operations (PR #233262)

[Matt Arsenault via llvm-branch-commits]

================
@@ -2172,6 +2172,19 @@ InstructionCost GCNTTIImpl::getMemoryOpCost(unsigned Opcode, Type *Src,
   bool IsDS = AddressSpace == AMDGPUAS::LOCAL_ADDRESS ||
               AddressSpace == AMDGPUAS::REGION_ADDRESS;
   bool HasDwordx3 = ST->hasDwordx3LoadStores();
+  bool HasUnalignedAccess = ST->hasUnalignedBufferAccessEnabled();
+  if (AddressSpace == AMDGPUAS::PRIVATE_ADDRESS ||
+      AddressSpace == AMDGPUAS::FLAT_ADDRESS)
+    HasUnalignedAccess = ST->hasUnalignedScratchAccessEnabled();
+  if (IsDS) {
----------------
arsenm wrote:

else 

https://github.com/llvm/llvm-project/pull/233262