[llvm] [AMDGPU] Reschedule loads in clauses to improve throughput (PR #102595)

[via llvm-commits]

================
@@ -242,6 +386,7 @@ bool SIPostRABundler::run(MachineFunction &MF) {
           BundleUsedRegUnits.reset();
         }
 
+        reorderLoads(MBB, BundleStart, Next);
----------------
ruiling wrote:

Instead of modifying BundleStart implicitly, would it be better to return the new start and assign to BundleStart after return from the call?

https://github.com/llvm/llvm-project/pull/102595