[PATCH] D15259: Normalize MBB's successors' probabilities in several locations.

[David Li via llvm-commits]

davidxl added a comment.

Are there any test case changes?


================
Comment at: lib/CodeGen/IfConversion.cpp:1263
@@ -1262,2 +1262,3 @@
     BBI.BB->addSuccessor(CvtBBI->FalseBB, NewFalse);
+    BBI.BB->normalizeSuccProbs();
   }
----------------
Add a brief comment about why normalization is needed (i.e., why the input BPs do not sum up to 1)?

================
Comment at: lib/CodeGen/IfConversion.cpp:1719
@@ -1717,3 +1718,3 @@
       // If the edge from ToBBI.BB to Succ already exists, update the
       // probability of this edge by adding NewWeight to it. An example is shown
       // below, in which A is ToBBI.BB and B is FromBBI.BB. In this case we
----------------
The comment needs to be fixed (probably in a different patch).

================
Comment at: lib/CodeGen/IfConversion.cpp:1753
@@ -1751,1 +1752,3 @@
 
+  ToBBI.BB->normalizeSuccProbs();
+  FromBBI.BB->normalizeSuccProbs();
----------------
It is unclear why this is needed.

================
Comment at: lib/CodeGen/IfConversion.cpp:1754
@@ +1753,3 @@
+  ToBBI.BB->normalizeSuccProbs();
+  FromBBI.BB->normalizeSuccProbs();
+
----------------
Will its successor edge be discarded later?


http://reviews.llvm.org/D15259