[llvm] Reapply "[SimplifyCFG] Fix profcheck failure from #213302" (#220932) (PR #220939)

[Yingwei Zheng via llvm-commits]

================
@@ -5117,6 +5115,56 @@ bool SimplifyCFGOpt::simplifySwitchOnSelectRemap(SwitchInst *SI,
       if (DTU && !OldDestStillTargeted)
         DTU->applyUpdates({{DominatorTree::Delete, BB, OldDest}});
     }
+
+    // Update the profile information on the switch if we had a profile
+    // for both it and the select instruction.
+    SmallVector<uint32_t> SwitchWeights;
+    bool SwitchHasBranchWeights = extractBranchWeights(*SI, SwitchWeights);
+    uint64_t SelectTrueWeight;
+    uint64_t SelectFalseWeight;
+    bool SelectHasBranchWeights =
+        extractBranchWeights(*Select, SelectTrueWeight, SelectFalseWeight);
+    if (Negate)
+      std::swap(SelectTrueWeight, SelectFalseWeight);
+    if (SwitchHasBranchWeights && SelectHasBranchWeights &&
+        !ProfcheckDisableMetadataFixes) {
+      uint64_t SwitchTotalWeight = sum_of(SwitchWeights, uint64_t{0});
+      SmallVector<uint64_t> NewSwitchWeights;
+      if (IsDefault) {
+        // Update the branch weights of the switch by multiplying all of them by
+        // the weight of the false branch of the select. Then add the new switch
+        // case at the end by multiplying the weight of the true branch of the
+        // select by the total weight of the switch.
+        NewSwitchWeights.reserve(SwitchWeights.size() + 1);
+        for (uint32_t SwitchWeight : SwitchWeights)
+          NewSwitchWeights.push_back(SwitchWeight * SelectFalseWeight);
----------------
dtcxzyw wrote:

This overestimates the branch weight for `x = K` because the original weight for `sel = K` also includes the contribution from `x = C`. We can correct this if we have an existing case for `sel = K`. But we cannot get the precise weight when `sel = K` falls into the default case.

Maybe it is not worth the complexity. But we can document the approximation here 


https://github.com/llvm/llvm-project/pull/220939