Re^2: Can you speed up method calls by eliminating method dispatch?

OK, thanks. I grok the aliasing, I think (some sort of NULL pointer changes to be a pointer to a CV, which new pointer holds the same address as Base::f's pointer?) With the Devel::Peek output, is this what's going on?

(Looking at P5 source, the real magic behind Devel::Peek seems to be in Perl_do_sv_dump() in dump.c.)

2 calls to Devel::Peek::Dump
Method call (dispatch)
1 call to Devel::Peek::Dump

Which translates to (first 2 calls):

SV = PVGV(0x816c6e8) at 0x814f6d8
  GP = 0x816c720
    CV = 0x814f714
SV = PVGV(0x816f938) at 0x814f810
  GP = 0x81709f8
    CV = 0x0             <-- &Child::f doesn't exist
    CVGEN = 0x0
[download]

(last):

SV = PVGV(0x816f938) at 0x814f810
  GP = 0x81709f8
    CV = 0x814f714       <-- &Child::f aliased to &Base::f
    CVGEN = 0x4d         <-- ...for method dispatch only
[download]

Within the pared-down Devel::Peek output:

SV = PVGV(0x816c6e8) at 0x814f6d8    # Reference (at 0x814f6d8) to typ
+eglob (at 0x816c6e8)?
  GP = 0x816c720                     # Address of typeglob? Has AV, HV
+, CV.
    CV = 0x814f714                   # Address of CV of f()

SV = PVGV(0x816f938) at 0x814f810    # Akin to above
  GP = 0x81709f8                     # ""
    CV = 0x0                         # Child::f()'s pointer to CV, equ
+iv place to Base::f()
    CVGEN = 0x0                      # "generated CV"? do you mean "th
+is CVGEN 'slot' is used for nothing else than method dispatch"?
[download]

Please ignore beyond here, massive tangent :)
No idea what 0x4d is, but within struct gp in gp.h, there is this:

U32 gp_cvgen; /* generational validity of cached gv_cv */

... and I can find stuff like this:

gv.c
453:                GvCVGEN(topgv) = PL_sub_generation;
[download]

... and PL_sub_generation seems to be incremented all over the place in the source. Furthermore:

gv.c
457-        else if (topgv && GvREFCNT(topgv) == 1) {
458-            /* cache the fact that the method is not defined */
459:            GvCVGEN(topgv) = PL_sub_generation;

...

1433-       /* Adding a new name to a subroutine invalidates method ca
+che */
1434:       PL_sub_generation++;
[download]

GvCVGEN is a macro that gets at the CVGEN in the Devel::Peek output:

dump.c
1544:   Perl_dump_indent(aTHX_ level, file, "    CVGEN = 0x%"UVxf"\n",
+ (UV)GvCVGEN(sv));
[download]

... so line 3330 would relate to your "&Child::f doesn't exist" above:

sv.c
3330-       GvCVGEN(dstr) = 0; /* Switch off cacheness. */
[download]

The zeroeth step in writing a module is to make sure that there isn't already a decent one in CPAN. (-- Pod::Simple::Subclassing)

Comment on Re^2: Can you speed up method calls by eliminating method dispatch? Select or Download Code

Replies are listed 'Best First'.
Re^3: Can you speed up method calls by eliminating method dispatch? by ikegami (Patriarch) on Jul 09, 2009 at 15:36 UTC
The point was that calling the method added the method to the child's namespace so that `@ISA` needs not be scanned on later calls to that method. Your questions go way beyond that, but I'll be happy to answer them anyway. I think (some sort of NULL pointer changes to be a pointer to a CV, which new pointer holds the same address as Base::f's pointer?) The `CV` field of the structure is the `CODE` slot of the glob. (See perldata) I was showing that `Child::f{CODE}` was assigned `Base::f{CODE}`. That means `Base::f` and `Child::f` are the same function at that point. do you mean "this CVGEN 'slot' is used for nothing else than method dispatch"? `Child::f{CODE}` is effectively undefined for everything that isn't a method call when `CVGEN` is true. No idea what `0x4d` is, but within struct gp in gp.h, there is [generational validity of cached gv_cv]* The cached entries need to be expired when there's a change to an `@ISA`, but keeping track of all the cached entries would require a fair amount of resource. Instead of clearing all the cached entries (which would require keeping track of them), changing any `@ISA` simply increments a counter. If the cache's counter (`CVGEN`) is lower than the change counter, the cache is stale and thus ignored. Or so goes my understanding.	[reply] [d/l] [select]
Re^4: Can you speed up method calls by eliminating method dispatch? by missingthepoint (Friar) on Jul 10, 2009 at 03:40 UTC
I grasped the main point - but my curiosity got the better of me :) Thanks for your detailed reply; it's highly appreciated. And your understanding (WRT cache expiration) makes perfect sense. The zeroeth step in writing a module is to make sure that there isn't already a decent one in CPAN. (-- Pod::Simple::Subclassing)	[reply]