Bug 74498 – ensure consistent heuristics

• Status: UNCONFIRMED
• Component: JavaScriptCore

[Andy Wingo]

Patch to be attached factors consistency-checking routine out of initializeOptions, née initializeHeuristics, and clamps various values at runtime.

Depends on bug 72938.

[Andy Wingo]

Created attachment 119203 [details]
Patch

[Eric Seidel (no email)]

It's best to CC the JSC guys if you want this reviewed. :)

[Martin Robinson]

Comment on attachment 119203 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=119203&action=review

> Source/JavaScriptCore/runtime/Options.cpp:51
> +#define CLAMP(var, min, max) var = ((var <= min) ? min : ((var >= max) ? max : var))

Perhaps it's possible to use clampTo from MathExtras.h here? It looks suspiciously like it already handles numeric_limits minimums, but I don't know if it has support for int32_t yet.

[Filip Pizlo]

Comment on attachment 119203 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=119203&action=review

> Source/JavaScriptCore/runtime/Options.cpp:55
> +    executionCounterValueForOptimizeNextInvocation = 0; // Not configurable.

That's not right. We want to be able to change this.

> Source/JavaScriptCore/runtime/Options.cpp:56
> +    CLAMP(executionCounterValueForOptimizeSoon, std::numeric_limits<int32_t>::min(), executionCounterValueForOptimizeNextInvocation - 1);

That's probably a bit too strict.  Could drop the - 1.

> Source/JavaScriptCore/runtime/Options.cpp:65
> +    CLAMP(desiredProfileFullnessRate, 0, desiredProfileLivenessRate);

Why?  These two nunbers are unrelated.

> Source/JavaScriptCore/runtime/Options.cpp:66
> +    CLAMP(numberOfGCMarkers, 1, 8);

Not sure about this. On Mac the default is to top out at 4 markers *by default*, but we want to enable more markers. What's wrong with having, say, 24 markers?  My machine has 24 logical CPUs.

> Source/JavaScriptCore/runtime/Options.cpp:82
> +    ASSERT((static_cast<int64_t>(executionCounterValueForOptimizeAfterLongWarmUp) << reoptimizationRetryCounterMax) >= static_cast<int64_t>(std::numeric_limits<int32_t>::min()));

I think that CheckedArithmetic.h might be better than this stuff.

> Source/JavaScriptCore/runtime/Options.h:68
> +    m(unsigned, reoptimizationRetryCounterMax,         30) \

Why set this to 30 and then recompute it anyway?