Speed up newSViv()

newSViv() is not used a whole lot by the core. But it is frequently
used in XS modules. In a nutshell, it allocates a new EMPTY SV just to
call sv_setiv which calls sv_upgrade which in turn spends inordinate
amounts of time looking at the properties of the SV to make it an
SVt_IV. But the properties of that SV are always the same: a clean
slate! Therefore, inlining the very simple bits of sv_setiv that are
actually necessary gives a very tangible speed-up.

It's not very easy to benchmark with with a language-level one-liner
because newSViv() isn't too common in the core. Thus follow XS
micro-benchmarks:

Benchmark 1: Virtually no-op XS function.

  SV *
  echo_integer(int in)
    CODE:
      RETVAL = newSViv(in);
    OUTPUT: RETVAL

  $ dumbbench -i50 --pin-frequency -- ./perl -Ilib \
    -MXS::APItest -e 'XS::APItest::echo_integer($_) for 1..1000000'

Before: 3.2782e-01 seconds
After:  3.0530e-01 seconds

A small change, but considering the massive overhead of a function call,
quite surprisingly noticeable.

Benchmark 2: XS function that constructs multiple integer values.

SV *
echo_integer_array(int in)
  PREINIT:
    int i;
    AV *av;
  CODE:
    av = newAV();
    RETVAL = newRV_noinc((SV *)av);
    av_extend(av, in-1);
    for (i = 0; i < in; ++i)
      av_store(av, i, newSViv(i));
  OUTPUT: RETVAL

  $ dumbbench -i50 --pin-frequency -- ./perl -Ilib \
    -MXS::APItest -e 'XS::APItest::echo_integer_array(100) for 1..10000'

Before: 1.18363e-01 seconds
After:  0.92050e-01 seconds

While in the grand scheme of things, this might seem like a very small
optimization, there are many XS modules that actually generate a lot of
integer values, de-serializers being good examples.

1 files changed, 16 insertions, 1 deletions

diff --git a/sv.c b/sv.c
index 4a2850e1f1..40b614e81e 100644
--- a/sv.c
+++ b/sv.c
@@ -9359,7 +9359,22 @@ Perl_newSViv(pTHX_ const IV i)
     SV *sv;
 
     new_SV(sv);
-    sv_setiv(sv,i);
+
+    /* Inlining ONLY the small relevant subset of sv_setiv here
+     * for performance. Makes a significant difference. */
+
+    /* We're starting from SVt_FIRST, so provided that's
+     * actual 0, we don't have to unset any SV type flags
+     * to promote to SVt_IV. */
+    assert(SVt_FIRST == 0);
+
+    SET_SVANY_FOR_BODYLESS_IV(sv);
+    SvFLAGS(sv) |= SVt_IV;
+    (void)SvIOK_on(sv);
+
+    SvIV_set(sv, i);
+    SvTAINT(sv);
+
     return sv;
 }