Bug 101667 – Improve effectiveness of function-level caching

• Status: RESOLVED
• Resolution: FIXED
• Component: New Bugs

[Oliver Hunt]

Improve effectiveness of function-level caching

[Oliver Hunt]

Created attachment 173150 [details]
Patch

[Oliver Hunt]

Committed r133975: <http://trac.webkit.org/changeset/133975>

[Darin Adler]

Comment on attachment 173150 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=173150&action=review

> Source/JavaScriptCore/runtime/CodeCache.h:55
> +template <typename KeyType, typename EntryType, int CacheSize> class Thingy {

Thingy? No, please use a real name. RandomEvictionCache maybe?

> Source/JavaScriptCore/runtime/CodeCache.h:59
> +    : m_randomGenerator((static_cast<uint32_t>(randomNumber() * UINT32_MAX)))

Incorrect indentation here.

> Source/JavaScriptCore/runtime/CodeCache.h:76
> +        size_t newIndex = m_randomGenerator.getUint32() % CacheSize;

This is not an even distribution across the cache; it will be slightly biased towards lower numbers unless the cache size is a power of 2. It would be better to make WeakRandom have a function to help you get even distribution.

[Oliver Hunt]

(In reply to comment #3)
> (From update of attachment 173150 [details])
> View in context: https://bugs.webkit.org/attachment.cgi?id=173150&action=review
> 
> > Source/JavaScriptCore/runtime/CodeCache.h:55
> > +template <typename KeyType, typename EntryType, int CacheSize> class Thingy {
> 
> Thingy? No, please use a real name. RandomEvictionCache maybe?

Doh!

> 
> > Source/JavaScriptCore/runtime/CodeCache.h:59
> > +    : m_randomGenerator((static_cast<uint32_t>(randomNumber() * UINT32_MAX)))
> 
> Incorrect indentation here.
Whoops
> 
> > Source/JavaScriptCore/runtime/CodeCache.h:76
> > +        size_t newIndex = m_randomGenerator.getUint32() % CacheSize;
> 
> This is not an even distribution across the cache; it will be slightly biased towards lower numbers unless the cache size is a power of 2. It would be better to make WeakRandom have a function to help you get even distribution.

The bias is not the worst thing in the world, adding the branches to get unbiased behaviour seems overkill.