Lexical subroutines can't capture refalias in the same block

[leonerd]

While investigating issues around refalias, declared_refs, and the upcoming PPC0034 (refalias in signatures), I encountered the following cornercase:

use v5.36;
use experimental qw( refaliasing declared_refs );

my \@arr = [1, 2, 3];
my sub inner { say "Inside inner, array is <@arr>"; }
say "Outside, array is <@arr>";
inner();

It produces output, showing that while the refalias affected the @arr variable in the outer scope, it did not affect the one captured by the inner sub:

Outside, array is <1 2 3>
Inside inner, array is <>

A slight modification of the program to put the lexical sub in its own nested block makes it behave as expected:

use v5.36;
use experimental qw( refaliasing declared_refs );

my \@arr = [1, 2, 3];
{
  my sub inner { say "Inside inner, array is <@arr>"; }
  say "Outside, array is <@arr>";
  inner();
}

Outside, array is <1 2 3>
Inside inner, array is <1 2 3>

[leonerd]

I believe the bug occurs due to the ordering of operations. If you use the -MO=Concise printer on the non-working case, you get the following output:

n  <@> leave[1 ref] vKP/REFC ->(end)
1     <0> enter v ->2
-     <@> lineseq vK ->4
2        <0> introcv[&inner:784,785] v ->3
3        <0> clonecv[&inner:784,785] v ->4
4     <;> nextstate(main 781 -:4) v:%,us,*,&,{,$,fea=15 ->5
a     <2> refassign[@arr:781,785] vK/LVINTRO,AV ->b
9        <@> anonlist sK*/1 ->a
5           <0> pushmark s ->6
6           <$> const[IV 1] s ->7
7           <$> const[IV 2] s ->8
8           <$> const[IV 3] s ->9
b     <;> nextstate(main 785 -:6) v:%,us,*,&,{,$,fea=15 ->c
i     <@> say vK ->j
c        <0> pushmark s ->d
h        <+> multiconcat("Outside, array is <>",19,1)[t7] sK/STRINGIFY ->i
-           <0> ex-pushmark s ->-
-           <0> ex-const s ->d
g           <@> join[t4] sK/2 ->h
d              <0> pushmark s ->e
-              <1> ex-rv2sv sK/STRICT,1 ->f
e                 <#> gvsv[*"] s ->f
f              <0> padav[@arr:781,785] l ->g
j     <;> nextstate(main 785 -:7) v:%,us,*,&,{,$,fea=15 ->k
m     <1> entersub vKS/STRICT ->n
-        <1> ex-list K ->m
k           <0> pushmark s ->l
-           <1> ex-rv2cv sK/STRICT,1 ->-
l              <0> padcv[&inner:784,785] s ->m

Notable is the fact that the clonecv operation (at node identified as 2) happened before the refassign operation (identified as a). This means the pad inside the inner CV captured the previous SV before the refassign happened.

Looking at the working second case, the refassign happens first, before the entire optree that represents that second block, and it is the second block that contains the clonecv operation. Due to this subsequent ordering, the clonecv can now see the correctly aliased SV, and the right behaviour results.

I believe therefore a fix for this bug (and a wider subset of the "refalias breaks closures" cornercases) may simply be to cause those introcv+clonecv operations to happen later in the outer optree.

Making this case work correctly should also cause similar cases using refalias in signatures as per PPC0034 to work correctly too.