Is not $_[0] aliased to the function argument? Why using all those references and dereferences?

Yes, it is aliased. But many people, including myself to some degree, have an aversion to using $_[n] in subroutines and will routinely change them to named vars if the subroutines become more complex than simple one liners.

Hence, when it is important to avoid copying scalar arguments, I prefer to use explicit references rather than implicit aliasing.

Moreover, running your code with N=2**29, I am still getting

I have no explanation for that. This is 5.18, with single and long excluded because they just take too long to run on strings of this length:

C:\test>\perl5.18\bin\perl.exe 1046579.pl -N=2**20
           Rate   classic translate       str
classic   482/s        --      -30%      -34%
translate 688/s       43%        --       -6%
str       730/s       52%        6%        --

C:\test>\perl5.18\bin\perl.exe 1046579.pl -N=2**29
             Rate   classic       str translate
classic   0.945/s        --      -21%      -26%
str        1.20/s       27%        --       -6%
translate  1.27/s       35%        6%        --

C:\test>\perl5.18\bin\perl.exe 1046579.pl -N=2**30
          s/iter   classic       str translate
classic     2.06        --      -19%      -26%
str         1.68       23%        --       -9%
translate   1.53       35%       10%        --

C:\test>\perl5.18\bin\perl.exe 1046579.pl -N=2**31
          s/iter       str   classic translate
str         3.97        --       -7%      -23%
classic     3.69        8%        --      -17%
translate   3.06       30%       20%        --
[download]

On my machine tr/// wins hands down for these longer lengths. Which is what I would expect everywhere.

It really surprises me that sub str does as well as it does for as long as it does given this complexity of opcodes from the explicit loop:

Compared the implicit loop of translate:

C:\test>perl -MO=Concise,translate 1046579.pl
main::translate:
8  <1> leavesub[1 ref] K/REFC,1 ->(end)
-     <@> lineseq KP ->8
1        <;> nextstate(main 1267 (eval 7):1) v:*,&,{,$ ->2
-        <1> null KS/2 ->-
6           <1> rv2sv sKRM/3 ->7
-              <@> scope sK ->6
-                 <0> ex-nextstate v ->2
5                 <2> aelem sKM/DREFSV,2 ->6
3                    <1> rv2av sKR/3 ->4
2                       <#> gv[*_] s ->3
4                    <$> const[IV 0] s ->5
7           <"> trans sS/COMPL ->8
1046579.pl syntax OK
[download]

And str() is still somewhat light on logic as it currently makes no attempt to deal with misalignments between the string length and the buffer size. I haven't checked to see what effect that has on the length of the string.

Why you would get different results to me for these large size strings; I cannot begin to guess?

With 2**30, it dies with Substitution loop at ./7.pl line 10.

I never saw that because I disabled that test along with single (for taking too long) before the error ever manifest itself.

I can confirm that I also get the error message once the length of the string moves above 2**29:

C:\test>\perl5.18\bin\perl.exe 1046579.pl -N=2**29
          s/iter      long   classic       str translate
long        3.92        --      -74%      -78%      -81%
classic     1.03      282%        --      -17%      -29%
str        0.847      363%       21%        --      -14%
translate  0.731      436%       40%       16%        --

C:\test>\perl5.18\bin\perl.exe 1046579.pl -N=2**30

Substitution loop at 1046579.pl line 42.
[download]

I have no idea what it means?