Hello,
without the patch (callgrind_annotate):


26,300,937  ???:_platform_strcmp [libsystem_platform.dylib]
13,288,306  Src/parse.c:ecstrcode [Src/zsh]
 4,486,910  Src/hashtable.c:hasher [Src/zsh]

after:

 2,661,459  ???:_platform_strcmp [libsystem_platform.dylib]
 2,588,524  Src/parse.c:ecstrcode [Src/zsh]
 5,959,019  Src/hashtable.c:hasher [Src/zsh]


So, for ecstrcode & hasher, the gain is 10.5 million, the loss is 1.5 million. Shell startup time measured with:

% repeat 10 { time /usr/local/bin/zsh-5.3.1-dev-0 -i -c exit } 

drops from 205 ms to 192 ms.

With additional has_token() patch, it's 205 ms to 190 ms. I think this can be assumed to be 15-20ms.

The 9 million gained shows that this is a serious optimization. Three more such optimizations, and gain would be 45-60 ms, a noticable difference.

Data: https://github.com/zdharma/hacking-private/tree/master/startup-dump-opt

Any ideas maybe, of what more can be done? I'm trying to add meaning to my functions-lexicon feature (automatically managed functions, loaded via autoload -w ...zwc). Dumps seem to be quite optimially mmapped, not sure where to look for something else related to Eprog.

--
Sebastian Gniazdowski
psprint /at/ zdharma.org