1. Drop default inline threshold back down to 200.
2. Do not use # of basic blocks as part of the cost computation since it doesn't really figure into function size.
3. More aggressively inline function with vector code.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@49061 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Transforms/Utils/InlineCost.cpp b/lib/Transforms/Utils/InlineCost.cpp
index 4349d0e..16f1264 100644
--- a/lib/Transforms/Utils/InlineCost.cpp
+++ b/lib/Transforms/Utils/InlineCost.cpp
@@ -234,14 +234,12 @@
// Now that we have considered all of the factors that make the call site more
// likely to be inlined, look at factors that make us not want to inline it.
- // Don't inline into something too big, which would make it bigger. Here, we
- // count each basic block as a single unit.
+ // Don't inline into something too big, which would make it bigger.
//
InlineCost += Caller->size()/20;
- // Look at the size of the callee. Each basic block counts as 20 units, and
- // each instruction counts as 5.
- InlineCost += CalleeFI.NumInsts*5 + CalleeFI.NumBlocks*20;
+ // Look at the size of the callee. Each instruction counts as 5.
+ InlineCost += CalleeFI.NumInsts*5;
return InlineCost;
}
@@ -258,9 +256,16 @@
if (CalleeFI.NumBlocks == 0)
CalleeFI.analyzeFunction(Callee);
+ float Factor = 1.0f;
+ // Single BB functions are often written to be inlined.
+ if (CalleeFI.NumBlocks == 1)
+ Factor += 0.5f;
+
// Be more aggressive if the function contains a good chunk (if it mades up
// at least 10% of the instructions) of vector instructions.
- if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)
- return 1.5f;
- return 1.0f;
+ if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/2)
+ Factor += 2.0f;
+ else if (CalleeFI.NumVectorInsts > CalleeFI.NumInsts/10)
+ Factor += 1.5f;
+ return Factor;
}