[llvm] [AMDGPU][True16] Select _t16 D16 loads/stores for FLAT/GLOBAL/SCRATCH (PR #210191)

[Jay Foad via llvm-commits]

================
@@ -3177,6 +3177,29 @@ void AMDGPUInstructionSelector::initM0(MachineInstr &I) const {
 bool AMDGPUInstructionSelector::selectG_LOAD_STORE_ATOMICRMW(
   MachineInstr &I) const {
   initM0(I);
+
+  // In true16 mode, a i8/i16 truncating store whose data operand is a
----------------
jayfoad wrote:

This feels like the wrong way to handle this. Instead, in true16 mode, we should not create truncating stores from i32->i16 and i32->i8 in the first place (but i16->i8 seems reasonable). See calls to legalForTypesWithMemDesc in AMDGPULegalizerInfo.

https://github.com/llvm/llvm-project/pull/210191