[Interpreter] Tune runtime profiler parameters for turbofan and OSR.

src/runtime-profiler.cc

Index: src/runtime-profiler.cc
diff --git a/src/runtime-profiler.cc b/src/runtime-profiler.cc
index ab76cc90649901d3f562de3f94c973ba1661e254..5cae78a75c4b9ee8b27a477189920a1f98ce7468 100644
--- a/src/runtime-profiler.cc
+++ b/src/runtime-profiler.cc
@@ -26,6 +26,9 @@ static const int kProfilerTicksBeforeBaseline = 1;
 // Number of times a function has to be seen on the stack before it is
 // optimized.
 static const int kProfilerTicksBeforeOptimization = 2;
+// Number of times a interpreted function has to be seen on the stack before
+// it is optimized (with Turbofan, ignition is only optimized with Turbofan).
+static const int kProfilerTicksBeforeOptimizingInterpretedFunction = 4;

[mythria, 2016/10/26 08:17:15]

In my experiments on Octane I found 4 to be a good value, we can tune it further
depending on the results from bots. 

 // If the function optimization was disabled due to high deoptimization count,
 // but the function is hot and has been seen on the stack this number of times,
 // then we try to reenable optimization for this function.
@@ -36,6 +39,7 @@ static const int kProfilerTicksBeforeReenablingOptimization = 250;
 static const int kTicksWhenNotEnoughTypeInfo = 100;
 // We only have one byte to store the number of ticks.
 STATIC_ASSERT(kProfilerTicksBeforeOptimization < 256);
+STATIC_ASSERT(kProfilerTicksBeforeOptimizingInterpretedFunction < 256);
 STATIC_ASSERT(kProfilerTicksBeforeReenablingOptimization < 256);
 STATIC_ASSERT(kTicksWhenNotEnoughTypeInfo < 256);
 
@@ -43,12 +47,12 @@ STATIC_ASSERT(kTicksWhenNotEnoughTypeInfo < 256);
 static const int kOSRCodeSizeAllowanceBase =
     100 * FullCodeGenerator::kCodeSizeMultiplier;

[mythria, 2016/10/26 08:17:15]

In my initial patch, I also changed this metric for default pipeline. It seems
to help Octane (+7% overall) but regresses sunspider (-1% overall) on the bots.
I can't reproduce the improvements on octane (Gameboy improves by 50% on the
bots), so I reset it to the default values. I will look into this after I
investigate why Gameboy and Box2d are varying a lot. 

 static const int kOSRCodeSizeAllowanceBaseIgnition =
-    100 * interpreter::Interpreter::kCodeSizeMultiplier;
+    10 * interpreter::Interpreter::kCodeSizeMultiplier;

[mythria, 2016/10/26 08:17:15]

This and the change to OSRCodeSizeAllowancePerTIck helps sunspider quite a bit.
Overall 11% improvement. Jetstream improves by 5%. If the bots are to be trusted
we get 1-2% overall improvement on Octane and Kraken. Though locally I don't see
similar improvements on octane. 

 
 static const int kOSRCodeSizeAllowancePerTick =
     4 * FullCodeGenerator::kCodeSizeMultiplier;
 static const int kOSRCodeSizeAllowancePerTickIgnition =
-    4 * interpreter::Interpreter::kCodeSizeMultiplier;
+    2 * interpreter::Interpreter::kCodeSizeMultiplier;
 
 // Maximum size in bytes of generated code for a function to be optimized
 // the very first time it is seen on the stack.
@@ -404,7 +408,7 @@ OptimizationReason RuntimeProfiler::ShouldOptimizeIgnition(
   SharedFunctionInfo* shared = function->shared();
   int ticks = shared->profiler_ticks();
 
-  if (ticks >= kProfilerTicksBeforeOptimization) {
+  if (ticks >= kProfilerTicksBeforeOptimizingInterpretedFunction) {
     int typeinfo, generic, total, type_percentage, generic_percentage;
     GetICCounts(function, &typeinfo, &generic, &total, &type_percentage,
                 &generic_percentage);