[llvm] [AMDGPU] Partially revert my llvm::less_second patch (PR #136615)

[Kazu Hirata via llvm-commits]

================
@@ -589,7 +589,7 @@ void PipelineSolver::populateReadyList(
   }
 
   if (UseCostHeur)
-    llvm::sort(ReadyList, llvm::less_second());
+    std::sort(ReadyList.begin(), ReadyList.end(), llvm::less_second());
----------------
kazutakahirata wrote:

> Does this mean that the code is gambling on the unstable sort order?

Probably.  We should probably use `llvm::stable_sort` or something, but for now I'm reverting the problematic part.  I've tried `llvm::stable_sort`, and that seems to work, too.  Shall we switch to `llvm::stable_sort`?


https://github.com/llvm/llvm-project/pull/136615