[lld] [llvm] [llvm][lld] Respect temporal profile weights in balanced partitioning (PR #213837)

[Karim Alweheshy via llvm-commits]

================
@@ -227,6 +241,14 @@ unsigned BalancedPartitioning::runIteration(const FunctionNodeRange Nodes,
       Signature.CachedGainLR = Cost - logCost(L - 1, R + 1);
     if (R > 0)
       Signature.CachedGainRL = Cost - logCost(L + 1, R - 1);
+    // Scaling a utility's move gain is mathematically equivalent to adding
+    // Weight distinct utility nodes with the same signature, but avoids
+    // expanding the graph. Weight == 1 deliberately takes no FP operation so
+    // unweighted inputs preserve their previous behavior.
+    if (Signature.Weight != 1) {
+      Signature.CachedGainLR *= static_cast<float>(Signature.Weight);
+      Signature.CachedGainRL *= static_cast<float>(Signature.Weight);
----------------
karim-alweheshy wrote:

Superseded by the sparse side-map design. Actual weighted runs multiply directly; absent or all-one maps dispatch to the original unweighted instantiation.

https://github.com/llvm/llvm-project/pull/213837