Fix B::Concise BEGIN skip count

[JRaspass]

B::Concise tries hard to hide its own BEGIN blocks when asked to report BEGIN blocks in end-user's code. It does this by maintaining a count of each BEGIN/use/no (and those of the dependencies) and then splicing these BEGIN blocks off before printing. However this count has drifted over time, which this commit fixes.

Before:

$ perl -MO=Concise,BEGIN,-exec -e 'BEGIN {1} BEGIN {2}'
... BEGIN 1-8 ...
BEGIN 9:
26 <;> nextstate(main 2 -e:1) v
27 <$> const[IV 1] s
28 <1> leavesub[1 ref] K/REFC,1
BEGIN 10:
29 <;> nextstate(main 4 -e:1) v
2a <$> const[IV 2] s
2b <1> leavesub[1 ref] K/REFC,1
-e syntax OK

After:

$ perl -MO=Concise,BEGIN,-exec -e 'BEGIN {1} BEGIN {2}'
BEGIN 1:
1  <;> nextstate(main 2 -e:1) v
2  <$> const[IV 1] s
3  <1> leavesub[1 ref] K/REFC,1
BEGIN 2:
4  <;> nextstate(main 4 -e:1) v
5  <$> const[IV 2] s
6  <1> leavesub[1 ref] K/REFC,1
-e syntax OK

The tests in ext/B/t/optree_specials.t now correctly only show the things being tested, but that also meant removing the bogus final test, as with this fix the code under test outputs nothing as it has no phase blocks in it.

• This set of changes requires a perldelta entry, and it is included.

[tonycoz]

I think it needs a perldelta entry, I can see people who've written code expecting the extra erroneous BEGINs being surprised (but hopefully not dismayed) by this.

As to what's in the perldelta entry, just add a new entry to Updated Modules and Errata matching the example pattern describing the change. It doesn't need to be detailed.

It looks like a reasonable change.

[JRaspass]

Amended to include a perldelta.

[leonerd]

LGTM I guess.

I agree this is an improvement on the prior broken state, but it still feels very fragile. I wonder if there's a more robust way this could be handled that didn't rely on it having to maintain a recursively-correct count that knows the number of all the BEGIN blocks it itself uses. Not to block this particular improvement, but perhaps something we can think about in future.

[tonycoz]

but it still feels very fragile.

When I review this I looked into whether we could automatically calculate the count at load time, but I think if B::Concise isn't loaded first that count is going to include other BEGIN blocks.

Though if B::Concise isn't loaded first we're probably going to strip the wrong BEGIN blocks, so I don't know.

Maybe we could add trivial BEGINs to the start and end of B::Concise and use those as brackets to select the BEGINs to eliminate.

[JRaspass]

Maybe we could add trivial BEGINs to the start and end of B::Concise and use those as brackets to select the BEGINs to eliminate.

@tonycoz that's what #23743 does, though I didn't consider code being loaded before B::Concise, I think I was mainly concerned with -MO= usage.