pp_ctl.c: one more unexercised exception in pp_goto()

[jkeenan]

This is the next ticket in a series of tickets originating in #23889 (which in turn originated in GH #23618). In preparation for the partial fatalization of goto LABEL, we've been examining the code in pp_ctl.c that governs the goto functionality. We've been trying to see that all exceptions defined in the relevant functions in pp_ctl.c are exercised ("covered") by the test suite.

In this ticket we turn our attention to this part of pp_ctl.c as it appears in blead as of v5.43.5-24-ge79b820fc6:

3364             if (!CvROOT(cv) && !CvXSUB(cv)) {
3365                 const GV * const gv = CvGV(cv);
3366                 if (gv) {
3367                     SV * const tmpstr = sv_newmortal();
3368                     gv_efullname3(tmpstr, gv, NULL);
3369                     DIE(aTHX_ "Goto undefined subroutine &%" SVf,  /* <-- case 'KKK' */
3370                                SVfARG(tmpstr));
3371                 }
3372                 DIE(aTHX_ "Goto undefined subroutine");  /* <-- case 'LLL' */
3373             }

(The KKK: and LLL: notations are from #23889 (comment).)

My research indicates that the KKK: exception is covered by the test suite in t/op/goto-sub.t here:

 68 # bug #99850, which is similar - freeing the subroutine we are about to
 69 # go(in)to during a FREETMPS call should not crash perl.
 70 
 71 package _99850 {
 72     sub reftype{}
 73     DESTROY { undef &reftype }
 74     eval { sub { my $guard = bless []; goto &reftype }->() };
 75 }
 76 like $@, qr/^Goto undefined subroutine &_99850::reftype at /,
 77    'goto &foo undefining &foo on sub cleanup';

This code was introduced, originally in t/op/goto.t, by Father C in 1d59c03 back in November 2011. The pp_ctl.c code, however, dates back much farther. It was introduced by @doughera in 4aa0a1f in June 1995:

4aa0a1f7324 (Andy Dougherty               1995-06-06 01:41:07 +0000 2103)                   DIE("Goto undefined subroutine &%s",SvPVX(tmpstr));
4aa0a1f7324 (Andy Dougherty               1995-06-06 01:41:07 +0000 2104)               }
4aa0a1f7324 (Andy Dougherty               1995-06-06 01:41:07 +0000 2105)               DIE("Goto undefined subroutine");

I don't believe the KKK: case was covered by the test suite until 2011. I can find no evidence that the LLL: case has ever been exercised by the test suite, and I have been pondering how it could be covered for the past week and have not come up with anything that works. Here are alternative ways we could deal with this situation:

• Someone could come up with a unit test that exercises line 3372. That test would go into t/op/goto-sub.t. Such a test would presumably cover the case where if (gv) at line 3366 evaluated false.

• We could decide line 3372 is absolutely unreachable; I'm not sure how we would revise pp_ctl.c in that case.

• In a somewhat similar case, @iabyn has suggested in pp_ctl.c: comment on unreachable code #23965 (review), we could hypothesize that though line 3372 is not exercisable by the test suite, someone could come up with some XS magic that exercises that line -- hence exception in that line should be retained, which conclusion we indicate with an inline comment.

Suggestions?

[iabyn]

On Sun, Nov 30, 2025 at 01:48:09PM -0800, James E Keenan wrote: 3364 if (!CvROOT(cv) && !CvXSUB(cv)) { 3365 const GV * const gv = CvGV(cv); 3366 if (gv) { 3367 SV * const tmpstr = sv_newmortal(); 3368 gv_efullname3(tmpstr, gv, NULL); 3369 DIE(aTHX_ "Goto undefined subroutine &%" SVf, /* <-- case 'KKK' */ 3370 SVfARG(tmpstr)); 3371 } 3372 DIE(aTHX_ "Goto undefined subroutine"); /* <-- case 'LLL' */ 3373 } Here are alternative ways we could deal with this situation: * Someone could come up with a unit test that exercises line 3372. That test would go into `t/op/goto-sub.t`.

I've been unable to think of a pure-Perl way of achieving this. It's unusual for pure-Perl subs not to have a name.

* We could decide line 3372 is absolutely unreachable;

It's definitely reachable via XS code which strips the name (GV) from a sub, a creates a sub without a name.

* In a somewhat similar case, @iabyn has suggested in #23965 (review), we could hypotheize that though line 3372 is not exercisable by the test suite, someone could come up with some XS magic that exercises that line -- hence exception in that line should be retained, which conclusion we indicate with an inline comment.

The code definitely needs to be kept. It's basically saying "if the CV has a name, include the name in the error message." If the CV hasn't got a name, then trying to print its name anyway will result in a SEGV.

[jkeenan]

Here are alternative ways we could deal with this situation:

• Someone could come up with a unit test that exercises line 3372. That
  test would go into t/op/goto-sub.t.

I've been unable to think of a pure-Perl way of achieving this. It's
unusual for pure-Perl subs not to have a name.
*>

• We could decide line 3372 is absolutely unreachable;

It's definitely reachable via XS code which strips the name (GV) from a
sub, a creates a sub without a name.
*>

• In a somewhat similar case, @iabyn has suggested in
  #23965 (review),
  we could hypotheize that though line 3372 is not exercisable by the test
  suite, someone could come up with some XS magic that exercises that line
  -- hence exception in that line should be retained, which conclusion we
  indicate with an inline comment.

The code definitely needs to be kept. It's basically saying "if the CV has
a name, include the name in the error message." If the CV hasn't got a
name, then trying to print its name anyway will result in a SEGV.

Okay, unless someone (hello, @tonycoz!) argues otherwise, I propose that we handle this the way we were originally going to handle a different exception in pp_goto(), namely, we'll insert an inline comment above line 3372 to the effect of "we can't reach this exception from our test suite, but could from XS code."

@iabyn, does the same logic apply in this case, also from inside the pp_goto() function?

3276             while (!CvROOT(cv) && !CvXSUB(cv)) {
3277                 const GV * const gv = CvGV(cv);
3278                 if (gv) {
3279                     GV *autogv;
3280                     SV *tmpstr;
3281                     /* autoloaded stub? */
3282                     if (cv != GvCV(gv) && (cv = GvCV(gv)))
3283                         continue;
3284                     autogv = gv_autoload_pvn(GvSTASH(gv), GvNAME(gv),
3285                                           GvNAMELEN(gv),
3286                                           GvNAMEUTF8(gv) ? SVf_UTF8 : 0);
3287                     if (autogv && (cv = GvCV(autogv)))
3288                         continue;
3289                     tmpstr = sv_newmortal();
3290                     gv_efullname3(tmpstr, gv, NULL);
3291                     DIE(aTHX_ "Goto undefined subroutine &%" SVf, SVfARG(tmpstr));
3292                 }
3293                 DIE(aTHX_ "Goto undefined subroutine"); /* EEE */
3294             }

Line 3293 (marked EEE) per earlier tickets is not covered by our test suite. (Line 3291 is covered -- twice -- in t/uni/goto.t.) Would line 3293 also be best handled by an inline comment? Thanks.

[tonycoz]

Okay, unless someone (hello, @tonycoz!) argues otherwise, I propose that we handle this the way we were originally going to handle a different exception in pp_goto(), namely, we'll insert an inline comment above line 3372 to the effect of "we can't reach this exception from our test suite, but could from XS code."

I don't see a way to do it either.

(The idea for the defer use of that code came from the checks of other call_sv() users when I worked on #16608.)

[iabyn]

On Mon, Dec 01, 2025 at 01:08:50PM -0800, James E Keenan wrote: @iabyn, does the same logic apply in this case, also from inside the `pp_goto()` function?

Yes.

[jkeenan]

On Mon, Dec 01, 2025 at 01:08:50PM -0800, James E Keenan wrote:
@iabyn, does the same logic apply in this case, also from inside the
pp_goto() function?
Yes.

Please see p.r. #23979. Thanks.