Optimize V8 memory layout of filter text

[mjethani]

Background

In EasyList+AA there are about 150 element hiding filters and exceptions of lengths exceeding 1,000 characters. Of these, about 20 are longer than 100,000 characters. Similarly, there are about 600 blocking and whitelist filters of lengths exceeding 1,000 characters, of which about 10 are longer than 100,000 characters. Most of the filter text is the list of domains, and this list is repeated often.

The memory layout of this text can be optimized by restructuring it in the following manner:

• Suppose there are two filters example.com#@#selector1 and example.com#@#selector2 occupying a total of 46 bytes
• First, we slice out the domain example.com (11 bytes) into its own string
• Next, we slice out the remainders #@#selector1 and #@#selector2 (12 bytes each)
• Next, we create two new concatenated strings example.com plus #@#selector1 and example.com plus #@#selector2 occupying a total of 35 bytes (11 plus 12 plus 12)

In order for this to work on V8, we would have to override some of V8's own string optimizations that are sometimes harmful (V8 issue ​#2869). When we "slice" out a part of a string, we must properly separate it from its V8 internal parent string, using an operation like JSON.parse(JSON.stringify(string)). Once we are able to do this, we can restructure such long filter texts with a more efficient layout in memory.

I have written a ​patch that performs this optimization on filters where the domains part of the text is 1,000 characters or longer. It reduces the initial heap used by ~3.8 MB, from ~39.6 MB to ~35.8 MB. More pertinently, it reduces the memory occupied by all strings from ~16 MB down to ~12 MB.

What to change

[...]

[sebastian]

Sorry, but we switched to GitLab. If this issue is still relevant, please file it again in the ​new issue tracker.