From: Nicholas Clark <nick@ccl4.org>
Date: Mon, 4 Jan 2010 13:03:31 +0000 (+0000)
Subject: A mechanism for inlineable OP equivalents of XSUBs is a TODO.
X-Git-Url: http://git.shadowcat.co.uk/gitweb/gitweb.cgi?a=commitdiff_plain;h=05fb4e2043bc02cebd4a05d37297970a1686d3b1;p=p5sagit%2Fp5-mst-13.2.git

A mechanism for inlineable OP equivalents of XSUBs is a TODO.
---

diff --git a/pod/perltodo.pod b/pod/perltodo.pod
index 4f8d77f..44638a8 100644
--- a/pod/perltodo.pod
+++ b/pod/perltodo.pod
@@ -624,6 +624,62 @@ bootstrap a cookbook.  (List::Util? Class::XSAccessor? Tree::Ternary_XS?)
 Another option could be deconstructing the implementation of some simpler
 functions in op.c.
 
+=head2 Allow XSUBs to inline themselves as OPs
+
+For a simple XSUB, often the subroutine dispatch takes more time than the
+XSUB itself. The tokeniser already has the ability to inline constant
+subroutines - it would be good to provide a way to inline other subroutines.
+
+Specifically, simplest approach looks to be to allow an XSUB to provide an
+alternative implementation of itself as a custom OP. A new flag bit in
+C<CvFLAGS()> would signal to the peephole optimiser to take an optree
+such as this:
+
+    b  <@> leave[1 ref] vKP/REFC ->(end)
+    1     <0> enter ->2
+    2     <;> nextstate(main 1 -e:1) v:{ ->3
+    a     <2> sassign vKS/2 ->b
+    8        <1> entersub[t2] sKS/TARG,1 ->9
+    -           <1> ex-list sK ->8
+    3              <0> pushmark s ->4
+    4              <$> const(IV 1) sM ->5
+    6              <1> rv2av[t1] lKM/1 ->7
+    5                 <$> gv(*a) s ->6
+    -              <1> ex-rv2cv sK ->-
+    7                 <$> gv(*x) s/EARLYCV ->8
+    -        <1> ex-rv2sv sKRM*/1 ->a
+    9           <$> gvsv(*b) s ->a
+
+perform the symbol table lookup of C<rv2cv> and C<gv(*x)>, locate the
+pointer to the custom OP that provides the direct implementation, and re-
+write the optree something like:
+
+    b  <@> leave[1 ref] vKP/REFC ->(end)
+    1     <0> enter ->2
+    2     <;> nextstate(main 1 -e:1) v:{ ->3
+    a     <2> sassign vKS/2 ->b
+    7        <1> custom_x -> 8
+    -           <1> ex-list sK ->7
+    3              <0> pushmark s ->4
+    4              <$> const(IV 1) sM ->5
+    6              <1> rv2av[t1] lKM/1 ->7
+    5                 <$> gv(*a) s ->6
+    -              <1> ex-rv2cv sK ->-
+    -                 <$> ex-gv(*x) s/EARLYCV ->7
+    -        <1> ex-rv2sv sKRM*/1 ->a
+    8           <$> gvsv(*b) s ->a
+
+I<i.e.> the C<gv(*)> OP has been nulled and spliced out of the execution
+path, and the C<entersub> OP has been replaced by the custom op.
+
+This approach should provide a measurable speed up to simple XSUBs inside
+tight loops. Initially one would have to write the OP alternative
+implementation by hand, but it's likely that this should be reasonably
+straightforward for the type of XSUB that would benefit the most. Longer
+term, once the run-time implementation is proven, it should be possible to
+progressively update ExtUtils::ParseXS to generate OP implementations for
+some XSUBs.
+
 =head2 Remove the use of SVs as temporaries in dump.c
 
 F<dump.c> contains debugging routines to dump out the contains of perl data