Add OP_MULTICONCAT op

Allow multiple OP_CONCAT, OP_CONST ops, plus optionally an OP_SASSIGN
or OP_STRINGIFY, to be combined into a single OP_MULTICONCAT op, which can
make things a *lot* faster: 4x or more.

In more detail: it will optimise into a single OP_MULTICONCAT, most
expressions of the form

    LHS RHS

where LHS is one of

    (empty)
    my $lexical =
    $lexical    =
    $lexical   .=
    expression  =
    expression .=

and RHS is one of

    (A . B . C . ...)            where A,B,C etc are expressions and/or
                                 string constants

    "aAbBc..."                   where a,A,b,B etc are expressions and/or
                                 string constants

    sprintf "..%s..%s..", A,B,.. where the format is a constant string
                                 containing only '%s' and '%%' elements,
                                 and A,B, etc are scalar expressions (so
                                 only a fixed, compile-time-known number of
                                 args: no arrays or list context function
                                 calls etc)

It doesn't optimise other forms, such as

    ($a . $b) . ($c. $d)

    ((($a .= $b) .= $c) .= $d);

(although sub-parts of those expressions might be converted to an
OP_MULTICONCAT). This is partly because it would be hard to maintain the
correct ordering of tie or overload calls.

The compiler uses heuristics to determine when to convert: in general,
expressions involving a single OP_CONCAT aren't converted, unless some
other saving can be made, for example if an OP_CONST can be eliminated, or
in the presence of 'my $x = .. ' which OP_MULTICONCAT can apply
OPpTARGET_MY to, but OP_CONST can't.

The multiconcat op is of type UNOP_AUX, with the op_aux structure directly
holding a pointer to a single constant char* string plus a list of segment
lengths. So for

    "a=$a b=$b\n";

the constant string is "a= b=\n", and the segment lengths are (2,3,1).
If the constant string has different non-utf8 and utf8 representations
(such as "\x80") then both variants are pre-computed and stored in the aux
struct, along with two sets of segment lengths.

For all the above LHS types, any SASSIGN op is optimised away. For a LHS
of '$lex=', '$lex.=' or 'my $lex=', the PADSV is optimised away too.

For example where $a and $b are lexical vars, this statement:

    my $c = "a=$a, b=$b\n";

formerly compiled to

    const[PV "a="] s
    padsv[$a:1,3] s
    concat[t4] sK/2
    const[PV ", b="] s
    concat[t5] sKS/2
    padsv[$b:1,3] s
    concat[t6] sKS/2
    const[PV "\n"] s
    concat[t7] sKS/2
    padsv[$c:2,3] sRM*/LVINTRO
    sassign vKS/2

and now compiles to:

    padsv[$a:1,3] s
    padsv[$b:1,3] s
    multiconcat("a=, b=\n",2,4,1)[$c:2,3] vK/LVINTRO,TARGMY,STRINGIFY

In terms of how much faster it is, this code:

    my $a = "the quick brown fox jumps over the lazy dog";
    my $b = "to be, or not to be; sorry, what was the question again?";

    for my $i (1..10_000_000) {
        my $c = "a=$a, b=$b\n";
    }

runs 2.7 times faster, and if you throw utf8 mixtures in it gets even
better. This loop runs 4 times faster:

    my $s;
    my $a = "ab\x{100}cde";
    my $b = "fghij";
    my $c = "\x{101}klmn";

    for my $i (1..10_000_000) {
        $s = "\x{100}wxyz";
        $s .= "foo=$a bar=$b baz=$c";
    }

The main ways in which OP_MULTICONCAT gains its speed are:

* any OP_CONSTs are eliminated, and the constant bits (already in the
  right encoding) are copied directly from the constant string attached to
  the op's aux structure.

* It optimises away any SASSIGN op, and possibly a PADSV op on the LHS, in
  all cases; OP_CONCAT only did this in very limited circumstances.

* Because it has a holistic view of the entire concatenation expression,
  it can do the whole thing in one efficient go, rather than creating and
  copying intermediate results. pp_multiconcat() goes to considerable
  efforts to avoid inefficiencies. For example it will only SvGROW() the
  target once, and to the exact size needed, no matter what mix of utf8
  and non-utf8 appear on the LHS and RHS.  It never allocates any
  temporary SVs except possibly in the case of tie or overloading.

* It does all its own appending and utf8 handling rather than calling
  out to functions like sv_catsv().

* It's very good at handling the LHS appearing on the RHS; for example in

    $x = "abcd";
    $x = "-$x-$x-";

  It will do roughly the equivalent of the following (where targ is $x);

    SvPV_force(targ);
    SvGROW(targ, 11);
    p = SvPVX(targ);
    Move(p,   p+1,  4, char);
    Copy("-", p,    1, char);
    Copy("-", p+5,  1, char);
    Copy(p+1, p+6,  4, char);
    Copy("-", p+10, 1, char);
    SvCUR(targ) = 11;
    p[11] = '\0';

  Formerly, pp_concat would have used multiple PADTMPs or temporary SVs to
  handle situations like that.

The code is quite big; both S_maybe_multiconcat() and pp_multiconcat()
(the main compile-time and runtime parts of the implementation) are over
700 lines each. It turns out that when you combine multiple ops, the
number of edge cases grows exponentially ;-)

1 files changed, 1 insertions, 0 deletions

diff --git a/pp_proto.h b/pp_proto.h
index e931546799..407cbd14a3 100644
--- a/pp_proto.h
+++ b/pp_proto.h
@@ -162,6 +162,7 @@ PERL_CALLCONV OP *Perl_pp_method_redir_super(pTHX);
 PERL_CALLCONV OP *Perl_pp_method_super(pTHX);
 PERL_CALLCONV OP *Perl_pp_mkdir(pTHX);
 PERL_CALLCONV OP *Perl_pp_modulo(pTHX);
+PERL_CALLCONV OP *Perl_pp_multiconcat(pTHX);
 PERL_CALLCONV OP *Perl_pp_multideref(pTHX);
 PERL_CALLCONV OP *Perl_pp_multiply(pTHX);
 PERL_CALLCONV OP *Perl_pp_nbit_and(pTHX);