Hello
(All test results and data are in:
https://github.com/zdharma/hacking-private/tree/master/eval-room )

I've compared time of 'source srctest.zsh' and 'eval "$(<srctest.zsh)" (the file has "a=1; b=2; c=3;" 300k times):

% time ./script1_srcopt.zsh  # source
./script1_srcopt.zsh  1.49s user 0.30s system 99% cpu 1.794 total

% time ./script2_srcopt.zsh  # eval
./script2_srcopt.zsh  1,01s user 0,08s system 98% cpu 1,105 total

This points that there should be much room for optimization?

Next I've run callgrind on each script, generated annotate, tweaked both results so that functions are 1-to-1 left and right. Diff generated with icdiff:

https://raw.githubusercontent.com/zdharma/hacking-private/master/eval-room/source_vs_eval.png

Some conclusions:

1. Difference is 0.9 * 10^9 instructions (eval has less).
2. Parser runs in almost *identical* way.
3. Biggest difference is in malloc-related functions:
   - szone_free_definite_size - 140 million,
   - szone_malloc_should_clear – 70 million,
   - szone_size – 83 million,
   - free – 40 million.
4. There's significant difference for memmove: 64 million.
5. For zhalloc it's 17 million.
6. Quite large difference for execlist: 27.8 million.
7. ingetc: 16.5 million.
8. strlen: 13.6 million.
9. For mutex functions, it's 52.5 million difference.

Epilogue, after Stephane catched the system-time use, we have 2x or more faster Zsh:

% time ./script1_srcopt.zsh # switched to zsh-5.3
./script1_srcopt.zsh  2,69s user 1,96s system 99% cpu 4,669 total

This is like a historical event IMO because frameworks can grow in size and/or stop being laggy.

Maybe the above results hint that something similar-easy is still possible? Or could POSIX getline() use be implemented (Joey Pabalinas offers a patch), to gain another milliseconds? It's pure gold for nothing.

PS. A plugin manager could cheat, set 0=... and do eval "$(<...)", but many plugins today use ${(%):-%N} instead of 0, and there's probably no way to cheat that?

--  
Sebastian Gniazdowski
psprint /at/ zdharma.org