He added

<#> gv[*b] s                      # my $b = $b[$i];
<1> rv2av sKR/1
<0> padsv[$i:2,4] s
<2> aelem sK/2
<0> padsv[$b:3,4] sRM*/LVINTRO
<2> sassign vKS/2
[download]

in order to replace a large number of

<#> gv[*b] s                      # $b[$i]
<1> rv2av sKR/1
<0> padsv[$i:2,3] s
<2> aelem sK/2
[download]

with

<0> padsv[$b:3,4] s               # $b
[download]

Assuming each of these ops take roughly the same amount of time to execute, this adds up. Basically,

my $b = $b[$i]; ... $b ... $b ... $b    # 6 + 1*3 = 9 ops
[download]

is faster than

$b[$i] ...  $b[$i]  ... $b[$i];         # 0 + 4*3 = 12 ops
[download]

Since his code looks up $b[i] 4 to 8 times (depending on the value of $b[$i]), this saves 6 to 18 ops. Per loop pass!