Lazily cache UTF-8 character length on RString

[lopopolo]

Store the UTF-8 length as computed by utf8_strlen on RString when MRB_UTF8_STRING is defined. This significantly (3200x) speeds up String operations at the cost of one mrb_int of memory per RString.

The UTF-8 char_len is computed on first call to utf8_strlen and is cached until mrb_str_modify is called, when it is zeroed out. utf8_strlen gets a new fastpath for when the character length is cached and retains the fast path for ASCII strings which is used as a fallback.

Benchmark

/usr/bin/time -l ./bin/mruby -e 'def str(len); ("💪" * len); end; s = str(100000); puts s.length; 1000000.times { s.length }'

This PR @ ca692e6

Completes in 0.30 seconds.

$ /usr/bin/time -l ./bin/mruby -e 'def str(len); ("💪" * len); end; s = str(100000); puts s.length; 1000000.times { s.length }'
100000
        0.30 real         0.30 user         0.00 sys
   2306048  maximum resident set size
         0  average shared memory size
         0  average unshared data size
         0  average unshared stack size
       572  page reclaims
         0  page faults
         0  swaps
         0  block input operations
         0  block output operations
         0  messages sent
         0  messages received
         0  signals received
         0  voluntary context switches
        48  involuntary context switches

master

Completes in 963 seconds. The caching implementation is 3210x faster than master.

$ /usr/bin/time -l ./bin/mruby -e 'def str(len); ("💪" * len); end; s = str(100000); puts s.length; 1000000.times { s.length }'
100000
      963.41 real       961.82 user         0.74 sys
   2252800  maximum resident set size
         0  average shared memory size
         0  average unshared data size
         0  average unshared stack size
       559  page reclaims
         0  page faults
         0  swaps
         0  block input operations
         0  block output operations
         0  messages sent
         0  messages received
         0  signals received
         0  voluntary context switches
    139399  involuntary context switches

[lopopolo]

Related to GH-4522 and GH-4529

cc @dearblue

[shuujii]

I think this patch will not be accepted as it is because the size of struct RString will increase (see: #3297).

However, it may be possible because the size does not increase in a limited environment such as 64-bit system + MRB_INT32 + boxing.

[shuujii]

64-bit system + MRB_INT32 + boxing

Sorry, boxing has nothing to do with it.

[lopopolo]

The tests do not currently pass when invoked like this:

CFLAGS="-Wall -DMRB_UTF8_STRING" rake test

str_replace does not work as expected. mruby-io seems to make assumptions about the layout of the RString struct. I think I bit off more than I can chew. Abandoning this PR.