JIT: Improve weight distribution in internal blocks produced by casts

src/coreclr/jit/morph.cpp

@@ -17697,6 +17697,9 @@ void Compiler::fgExpandQmarkForCastInstOf(BasicBlock* block, Statement* stmt)
 
     assert(qmark->gtFlags & GTF_QMARK_CAST_INSTOF);
 
+    const BasicBlock::weight_t currBbWeight = block->bbWeight;
+    const BasicBlock::weight_t nextBbWeight = (block->bbNext != nullptr) ? block->bbNext->bbWeight : currBbWeight;
+
     // Get cond, true, false exprs for the qmark.
     GenTree* condExpr  = qmark->gtGetOp1();
     GenTree* trueExpr  = qmark->gtGetOp2()->AsColon()->ThenNode();
@@ -17782,11 +17785,26 @@ void Compiler::fgExpandQmarkForCastInstOf(BasicBlock* block, Statement* stmt)
     cond1Block->bbJumpDest = remainderBlock;
     cond2Block->bbJumpDest = remainderBlock;
 
-    // Set the weights; some are guesses.
+    // Currently, we don't instrument internal blocks, so the only way we can set weights to these blocks
+    // is to analyze successors and take a guess.
+    BasicBlock::weight_t castSuccessLikelihood = BB_ZERO_WEIGHT;
+
+    // We don't expand casts inside rarely executed blocks, but currently we skip only blocks
+    // with BBF_RUN_RARELY flag and still can hit zero weight here (in that cases all internal blocks
+    // will also have zero weight).
+    if (currBbWeight > BB_ZERO_WEIGHT)
+    {
+        castSuccessLikelihood = clamp(nextBbWeight / currBbWeight, BB_ZERO_WEIGHT, 1.0f);
+    }
+
     asgBlock->inheritWeight(block);
     cond1Block->inheritWeight(block);
-    cond2Block->inheritWeightPercentage(cond1Block, 50);
-    helperBlock->inheritWeightPercentage(cond2Block, 50);
+
+    // cond2Block is always taken if the cast always succeeds (helperBlock will be cold in this case)
+    // If it always fails the only guess we can make that it's either object is null or of a
+    // wrong type (50/50).
+    cond2Block->inheritWeightPercentage(block, (UINT32)(castSuccessLikelihood * 50.0f) + 50);
+    helperBlock->inheritWeightPercentage(block, 50 - (UINT32)(castSuccessLikelihood * 50.0f));
 
     // Append cond1 as JTRUE to cond1Block
     GenTree*   jmpTree = gtNewOperNode(GT_JTRUE, TYP_VOID, condExpr);

src/coreclr/jit/utils.h

@@ -41,6 +41,13 @@ inline bool isPow2(T i)
     return (i > 0 && ((i - 1) & i) == 0);
 }
 
+// Clamps the given value between the given lower and upper values
+template <typename T>
+inline bool clamp(T value, T lower, T upper)
+{
+    return max(lower, min(value, upper));
+}
+
 // Adapter for iterators to a type that is compatible with C++11
 // range-based for loops.
 template <typename TIterator>