You may be better off doing the initial check without using the regex engine; only using it with split where necessary.

As you can see from ++aaron_baugher's analysis, your results will depend on your real data. Furthermore, if your volume of data is small, your choice of solution may make little difference (in terms of runtime).

Here's a solution using substr, rindex and length for the initial check. As a proof-of-concept to show that these functions work on characters (as opposed to bytes), I've included single-byte and multi-byte characters in the data.

#!/usr/bin/env perl -l

use strict;
use warnings;
use utf8;
use open OUT => ':utf8', ':std';

my @strings = (
    'A B C', 'D E', 'F ', 'G', 'H  ', 'I   ',
    "\N{MERCURY} \N{FEMALE SIGN} \N{EARTH}", "\N{MALE SIGN} \N{JUPITER
+}",
    "\N{SATURN} ", "\N{URANUS}", "\N{NEPTUNE}  ", "\N{PLUTO}   ",
);

for (@strings) {
    next if substr($_, -1, 1) eq ' ' or rindex($_, ' ', length() - 2) 
+== -1;
    print;
}
[download]

Output:

A B C
D E
☿ ♀ ♁
♂ ♃

[The Unicode range of characters labelled "Astrological symbols" is 0x263d to 0x2647. There is no charname for "VENUS" or "MARS"; the charnames "FEMALE SIGN" and "MALE SIGN" are defined for these symbols, respectively.]

Here's a benchmark test. This uses my sample data. If you choose this route, you should benchmark with representative samples of your data.

#!/usr/bin/env perl -l

use strict;
use warnings;

use Benchmark qw{cmpthese};

my @strings = (
    'A B C', 'D E', 'F ', 'G', 'H  ', 'I   ',
    "\N{MERCURY} \N{FEMALE SIGN} \N{EARTH}", "\N{MALE SIGN} \N{JUPITER
+}",
    "\N{SATURN} ", "\N{URANUS}", "\N{NEPTUNE}  ", "\N{PLUTO}   ",
);

my $re = qr{\s+\b};

cmpthese -1 => {
    no_re_check_and_split => \&no_re_check_and_split,
    re_check_and_split => \&re_check_and_split,
    split_and_re_check => \&split_and_re_check,
};

sub no_re_check_and_split {
    for (@strings) {
        next if substr($_, -1, 1) eq ' ' or rindex($_, ' ', length() -
+ 2) == -1;
        my @parts = split /$re/;
    }
}

sub re_check_and_split {
    for (@strings) {
        next unless /$re/;
        my @parts = split /$re/;
    }
}

sub split_and_re_check {
    for (@strings) {
        my @parts = split /$re/;
        next if @parts > 1;
    }
}
[download]

Here's three sample runs:

                         Rate split_and_re_check re_check_and_split no
+_re_check_and_split
split_and_re_check    50243/s                 --               -18%   
+               -29%
re_check_and_split    61265/s                22%                 --   
+               -13%
no_re_check_and_split 70274/s                40%                15%   
+                 --

                         Rate split_and_re_check re_check_and_split no
+_re_check_and_split
split_and_re_check    53593/s                 --               -19%   
+               -27%
re_check_and_split    66370/s                24%                 --   
+               -10%
no_re_check_and_split 73770/s                38%                11%   
+                 --

                         Rate split_and_re_check re_check_and_split no
+_re_check_and_split
split_and_re_check    53096/s                 --               -20%   
+               -27%
re_check_and_split    66369/s                25%                 --   
+                -8%
no_re_check_and_split 72404/s                36%                 9%   
+                 --
ken@ganymede: ~/tmp
[download]

With my sample data, doing the initial check without a regex appears faster. Again, I'll stress, you'll need to check with your data.