Project

General

Profile

Misc #15806

Explicitly initialise encodings on init to remove branches on encoding lookup

Added by methodmissing (Lourens Naudé) 3 months ago.

Status:
Open
Priority:
Normal
Assignee:
-
[ruby-core:92452]

Description

References Github PR https://github.com/ruby/ruby/pull/2128

I noticed that the encoding table is loaded on startup of even just miniruby (minimal viable interpreter use case) through this backtrace during ruby setup:

/home/lourens/src/ruby/ruby/miniruby(rb_enc_init+0x12) [0x56197b0c0c72] encoding.c:587
/home/lourens/src/ruby/ruby/miniruby(rb_usascii_encoding+0x1a) [0x56197b0c948a] encoding.c:1357
/home/lourens/src/ruby/ruby/miniruby(Init_sym+0x7a) [0x56197b24810a] symbol.c:42
/home/lourens/src/ruby/ruby/miniruby(rb_call_inits+0x1d) [0x56197b11afed] inits.c:25
/home/lourens/src/ruby/ruby/miniruby(ruby_setup+0xf6) [0x56197b0ec9d6] eval.c:74
/home/lourens/src/ruby/ruby/miniruby(ruby_init+0x9) [0x56197b0eca39] eval.c:91
/home/lourens/src/ruby/ruby/miniruby(main+0x5a) [0x56197b051a2a] ./main.c:41

Therefore I think it makes sense to instead initialize encodings explicitly just prior to symbol init, which is the first entry point into the interpreter loading that currently triggers rb_enc_init and remove the initialization check branches from the various lookup methods.

Some of the branches collapsed, cachegrind output, columns are Ir Bc Bcm Bi Bim with Ir (instructions retired), Bc (branches taken) and Bcm (branches missed) relevant here as there are no indirect branches (function pointers etc.):

(hot function, many instructions retired and branches taken and missed)

         .          .       .          .       .  rb_encoding *
         .          .       .          .       .  rb_enc_from_index(int index)
   835,669          0       0          0       0  {
13,133,536  6,337,652  50,267          0       0      if (!enc_table.list) {
         3          0       0          0       0    rb_enc_init();
         .          .       .          .       .      }
23,499,349  8,006,202 293,161          0       0      if (index < 0 || enc_table.count <= (index &= ENC_INDEX_MASK)) {
         .          .       .          .       .    return 0;
         .          .       .          .       .      }
30,024,494          0       0          0       0      return enc_table.list[index].enc;
 1,671,338          0       0          0       0  }

(cold function, representative of the utf8 variant more or less too)

         .          .       .          .       .  rb_encoding *
         .          .       .          .       .  rb_ascii8bit_encoding(void)
         .          .       .          .       .  {
    27,702      9,235     955          0       0      if (!enc_table.list) {
         .          .       .          .       .    rb_enc_init();
         .          .       .          .       .      }
     9,238          0       0          0       0      return enc_table.list[ENCINDEX_ASCII].enc;
     9,232          0       0          0       0  }

I think lazy loading encodings and populating the table is fine, but initializing it can be done more explicitly in the boot process.

Also available in: Atom PDF