mruby-regexp: set $&, $`, $' and $+ after a match - #7025

create_matchdata() assigns $~ and $1 to $9, but leaves the four remaining
match globals untouched. They read as nil after a match that already holds
everything they describe.

/b(c)/ =~ "abc"
$&    # CRuby: "bc",  mruby: nil
$`    # CRuby: "a",   mruby: nil
$'    # CRuby: "",    mruby: nil
$+    # CRuby: "c",   mruby: nil

Cause

Nothing below regexp.c treats these names specially.
PM_BACK_REFERENCE_READ_NODE in codegen.c emits a plain OP_GETGV on the
reference's own name, the same opcode PM_GLOBAL_VARIABLE_READ_NODE emits, so a
read of $& resolves through the ordinary global table just as $1 does. The
compiler side already works; the four are simply never assigned.

Change

Assign them in create_matchdata(), next to the $1 to $9 loop, from the same
capture array and through the same re_byte_substr() helper. That is the only
point at which a successful match is known together with both the subject and the
capture array, so its four call sites (exec_match(), regexp_gsub_str(),
regexp_sub_str() and regexp_scan()) inherit the assignment exactly as they
inherit $~ today.

$&, $` and $' follow from the whole match offsets. The latter two are
the expressions matchdata_pre() and matchdata_post() evaluate, but those are
instance methods that need a receiver, so routing the globals through them would
mean a method dispatch from C for what one re_byte_substr() call does.

$+ is the last group that actually participated, which is not
captures[(num_captures-1)*2], because the last group in the pattern is
frequently the one that did not match:

/(a)|(b)/ =~ "a"
$+    # "a", from group 1, since group 2 never participated

so it needs a backwards scan from md->num_captures - 1 down to group 1. A
pattern with no groups leaves that scan with nothing to find and yields nil,
which is what CRuby reports.

clear_match_globals() sets the four to nil alongside $~ and the nth_syms
loop. Without that a failed match leaves a stale $& visible, which is the same
defect in the other direction and is what the existing test for $1 guards
against.

ensure_nth_syms() interns the four and is renamed to ensure_match_syms(), now
that it covers more than $1 to $9. Both callers already invoke it, so no new
call site is needed and the interning stays out of the per-match path.

regexp_match_p() deliberately sets no match globals, and the existing
Regexp#match? - does not update last match test asserts that. It keeps setting
none.

Cost and lifetime

re_byte_substr() allocates through mrb_str_new() and therefore copies the
bytes, so the four values describe the subject as it was at match time and no
later mutation of it can invalidate them.

The cost is four mrb_gv_set() calls per match on top of the nine the $1 to
$9 loop already issues, and four more strings retained until the next match or
the next clear. $` and $' are the ones that matter there, since between
them they can pin a copy of nearly the whole subject. Globals are a direct GC
root, so mrb_gv_set() needs no write barrier and none was added.

In regexp_gsub_str() and regexp_scan() the loop's mrb_gc_arena_restore()
runs before the trailing create_matchdata() call, so the new allocations are
made after the arena is already unwound.

Known limit

Assigning $~ directly still updates none of the four, because they are set only
from inside create_matchdata(). That is the existing behaviour of $1 to $9,
which these four now share rather than diverge from; closing it would mean
deriving all of them from $~ at read time, which is a separate change.

Testing

rake test passes: 1950 tests, 0 failures, 0 crashes. It also passes with
MRB_UTF8_STRING enabled (1970 tests, 0 failures), where the four report the
same multibyte substrings as CRuby:

$ cat t.rb
"日本語です" =~ /本(語)/
p [$&, $`, $', $+]
$ ./build/utf8/bin/mruby t.rb
["本語", "日", "です", "語"]

Behaviour checked against CRuby 4.0.6.

Summary by CodeRabbit

• New Features

  • Regular expression matches now provide whole-match, pre-match, post-match, and last-capture values through the corresponding global match variables.
  • These values are cleared when a match fails.

• Bug Fixes

  • Corrected handling of global capture variables, including cases with optional or missing captures.