[llvm] [AMDGPU] Classify FLAT instructions as VMEM (PR #137148)

[Matt Arsenault via llvm-commits]

================
@@ -2454,8 +2454,8 @@ bool SIInsertWaitcnts::isPreheaderToFlush(
 }
 
 bool SIInsertWaitcnts::isVMEMOrFlatVMEM(const MachineInstr &MI) const {
-  return SIInstrInfo::isVMEM(MI) ||
-         (SIInstrInfo::isFLAT(MI) && mayAccessVMEMThroughFlat(MI));
+  return (SIInstrInfo::isFLAT(MI) && mayAccessVMEMThroughFlat(MI)) ||
+         SIInstrInfo::isVMEM(MI);
----------------
arsenm wrote:

This amounts to just isVMEM. I'm not sure why this is bothering to specially handle the case where a flat instruction is statically known to only access LDS. I think the only way that would happen is if we had a volatile flat access to an LDS variable, which is mildly useful. 

This should probably be something like if (isFLAT()) return mayAccessVMEMThroughFlat(); else // other non-flat cases 


https://github.com/llvm/llvm-project/pull/137148