split is faster, but also less versatile:

$ perl -MBenchmark=cmpthese -wE'my$x=join"    "=>"",("aaa")x5,"";say"s
+ourc: |$x|";sub splt{split" ",$x,1};sub rgx{$x=~s/^\s+//r};say "split
+: |",splt(),"|";say"regex: |",rgx(),"|";cmpthese(-2,{splt=>\&splt,rgx
+=>\&rgx})'
sourc: |    aaa    aaa    aaa    aaa    aaa    |
split: |aaa    aaa    aaa    aaa    aaa    |
regex: |aaa    aaa    aaa    aaa    aaa    |
          Rate  rgx splt
rgx  3382502/s   -- -58%
splt 7989872/s 136%   --
[download]

and split is not able to trim trailing space without help

$ perl -MBenchmark=cmpthese -wE'my$x=join"    "=>"",("abc")x5,"";say"s
+ourc: |$x|";sub splt{split" ",reverse((split" ",(reverse$x),1)[0]),1;
+};sub rgx{$x=~s/^\s+//r=~s/\s+$//r};say "split: |",splt(),"|";say"reg
+ex: |",rgx(),"|";cmpthese(-2,{splt=>\&splt,rgx=>\&rgx})'
sourc: |    abc    abc    abc    abc    abc    |
split: |abc    abc    abc    abc    abc|
regex: |abc    abc    abc    abc    abc|
          Rate  rgx splt
rgx  1047602/s   -- -64%
splt 2915089/s 178%   --
[download]

It is still faster, but I would pity the coder that has to maintain it.

So, concluding split (as a function) is faster, the next step is to profile your complete process and wonder if winning this little is worth the maintainability nightmare.

If this functionality only uses 0.01% of the complete time, I would not even think about bothering.

If it makes up 75% of the process time, and you are dealing with 4 Tb datastreams I would create an XS module that does exactly what I need and squeeze every CPU microsecond out of that optimized code.

I seriously doubt if that would be worth it. Seriously. There are millions of existing scripts that already use s{^\s+}{} and s{\s+$}{} (which removes all whitespace, so my functions may not be identical for some data), and people that use Perl, are familiar with it. Introducing functions like ltrim and rtrim or whatever they are to be called will for sure appeal to many users (unless the syntax is not obvious: as an excercise to the reader, please try to build split from scratch and implement all edge cases).

Is presented as an XS module, it may be used as playground, but then one has to ask themselves, does the time required to import the module weigh against the time gained in the function?