Seems to be the case that named hash lookups are (marginally, but consistently) quicker:

{ 
    my $ref = { 1 .. 1e6 }; 
    my %h = 1 .. 1e6; 
    cmpthese -5,{
        a => sub{ exists $ref->{$_} and 1 for 1 .. 1e6; },
        b => sub{ exists $h{$_}     and 1 for 1 .. 1e6; }
    }
};;
    Rate   a   b
a 1.24/s  -- -3%
b 1.28/s  3%  --

    Rate   a   b
a 1.36/s  -- -5%
b 1.43/s  5%  --

    Rate   a   b
a 1.24/s  -- -3%
b 1.28/s  3%  --

    Rate   a   b
a 1.23/s  -- -5%
b 1.29/s  5%  --

    Rate   a   b
a 1.27/s  -- -3%
b 1.30/s  3%  --
[download]

The same seems to be true for package variables ( scalar holding hashref and hash in the symbol table), and my variable lookup seems to be faster:

our $ref = { 1 .. 1e6 }; 
our %h = 1 .. 1e6; 
{ 
    my $_ref = { 1 .. 1e6 }; 
    my %_h = 1 .. 1e6; 
    cmpthese -5,{
        a => sub{ exists $_ref->{$_} and 1 for 1 .. 1e6; },
        b => sub{ exists $_h{$_}     and 1 for 1 .. 1e6; },
        c => sub{ exists $ref->{$_} and 1 for 1 .. 1e6; },
        d => sub{ exists $h{$_}     and 1 for 1 .. 1e6; }
    }
};;
    Rate   a   c   d   b
a 3.31/s  -- -2% -3% -4%
c 3.38/s  2%  -- -0% -2%
d 3.39/s  3%  0%  -- -2%
b 3.45/s  4%  2%  2%  --
[download]

Thanks for measurement and correction.

perl -le'print map{pack c,($-++?1:13)+ord}split//,ESEL'

Maybe the verification of the reference type explains the difference?
لսႽ† ᥲᥒ⚪⟊Ⴙᘓᖇ Ꮅᘓᖇ⎱ Ⴙᥲ𝇋ƙᘓᖇ

The difference appears to be that a pad lookup requires a single opcode where dereferencing a hashref requires three opcodes:

9     <;> nextstate(main 47 -e:1) v:%,{ ->a                   f     <;
+> nextstate(main 47 -e:1) v:%,{ ->g
a     <;> nextstate(main 47 -e:1) v:%,{ ->b                   g     <;
+> nextstate(main 47 -e:1) v:%,{ ->h
m     <2> leaveloop vK/2 ->n                                  u     <2
+> leaveloop vK/2 ->v
f        <{> enteriter(next->j last->m redo->g) lKS/8 ->k     l       
+ <{> enteriter(next->r last->u redo->m) lKS/8 ->s
-           <0> ex-pushmark s ->b                             -       
+    <0> ex-pushmark s ->h
-           <1> ex-list lK ->e                                -       
+    <1> ex-list lK ->k
b              <0> pushmark s ->c                             h       
+       <0> pushmark s ->i
c              <$> const[IV 1] s ->d                          i       
+       <$> const[IV 1] s ->j
d              <$> const[IV 10] s ->e                         j       
+       <$> const[IV 10] s ->k
e           <#> gv[*_] s ->f                                  k       
+    <#> gv[*_] s ->l
-        <1> null vK/1 ->m                                    -       
+ <1> null vK/1 ->u
l           <|> and(other->g) vK/1 ->m                        t       
+    <|> and(other->m) vK/1 ->u
k              <0> iter s ->l                                 s       
+       <0> iter s ->t
-              <@> lineseq vK ->-                             -       
+       <@> lineseq vK ->-
i                 <1> exists vK/1 ->j                         q       
+          <1> exists vK/1 ->r
-                    <1> ex-helem sK/2 ->i                    -       
+             <1> ex-helem sK/2 ->q
g                       <0> padhv[%h:46,47] sR ->h            o       
+                <1> rv2hv[t8] sKR/1 ->p
                                                              n       
+                   <1> rv2sv sKM/DREFHV,1 ->o
                                                              m       
+                      <#> gv[*r] s ->n
-                       <1> ex-rv2sv sK/1 ->-                 -       
+                <1> ex-rv2sv sK/1 ->-
h                          <#> gvsv[*_] s ->i                 p       
+                   <#> gvsv[*_] s ->q
j                 <0> unstack v ->k                           r       
+          <0> unstack v ->s
-e syntax OK                                                  -e synta
+x OK
[download]

---

With the rise and rise of 'Social' network sites: 'Computers are making people easier to use everyday'

Examine what is said, not who speaks -- Silence betokens consent -- Love the truth but pardon error.

"Science is about questioning the status quo. Questioning authority".

In the absence of evidence, opinion is indistinguishable from prejudice.

The reference is a named variable though, so that also needs a lookup in the symbol table or pad, plus needs to then be dereferenced.