According to perlreguts, the RE engine also uses fbm_index() to scan for the leftmost atom. There shouldn't be any reason why the performance of the two should differ by a large margin, and I would expect the regular expression to be a bit slower in the general case due to the setup. So I think it's either that your data somehow favours a branch in the RE engine that goes to fbm_index faster, or that the benchmark is not measuring what you want. But I also vaguely remember some thread about such a discrepancy on this site, but I can't find it is index faster than regexp for fixed text token?.

git blame tells me nobody touched index since 2009, and that change was some refcounting change. The other changes were in 2006.

If you don't mind, could you post the benchmark code? I'm just curious to look it over.

Hi davido

well it was in June, below what I found on my disk:

There is an "X" in the middle of a 52MB string of repeated alphabet letters ($pattern = join "","a".."z")

Depending if you look for "Xabc..xyz" or "abc...xyzX" the different approaches show their strength.

I did more tests which I can't find anymore strongly indicating that index doesn't use Boyer-Moore.

Just vary the position of the "X". I'd be glad if you looked it over. :)

use Time::HiRes qw[ time ];
 
 
my $pattern = join "","a".."z";
 
my $str= $pattern x 1E6 . "X" .$pattern x1E6;
$\="\n";
$|=1;
 
print "Length: ",length $str;
 
print "\n---End X";
 
$start=time;
print "Match: ", $str =~/${pattern}X/;
printf "\t took %.3f sec\n",time-$start;
 
$start=time;
print "Index: ", index $str , "${pattern}X";
printf "\t took %.3f sec\n",time-$start;
 
print "\n---Start X";
 
$start=time;
print "Match: ", $str =~/X${pattern}/;
printf "\t took %.3f sec\n",time-$start;
 
$start=time;
print "Index: ", index $str , "X${pattern}";
printf "\t took %.3f sec\n",time-$start;
[download]

RESULT:

Length: 52000001
 
---End X
Match: 1
     took 0.021 sec
Index: 25999974
     took 0.263 sec
 
---Start X
Match: 1
     took 0.165 sec
Index: 26000000
     took 0.094 sec
[download]

Cheers Rolf

It's not that index doesn't use Boyer Moore, it's more that the regex engine doesn't always use Boyer Moore. If the regex engine comes to the conclusion to use screaminstr, it will use that instead of Boyer Moore.

I guess that if you change your data to be more favorable to Boyer Moore, the results will reverse as the regular expression will still not use Boyer Moore.

Update: Even weirder. Looking at the log of use re 'Debug','ALL';, it claims that fbm_instr() is used:

C:\Projekte>perl -Mre=Debug,ALL -e "shift=~/abcdefghijklmnopqrstuvwxyz
+X/" abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabcdef
ghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzXabcdefghijklmnopqrstuvw
+xyzabcdefghijklmnopqrstuvwxyz
Compiling REx "abcdefghijklmnopqrstuvwxyzX"
Starting first pass (sizing)
 >abcdefghij... |   1|  reg
                |    |    brnc
                |    |      piec
                |    |        atom
Required size 9 nodes
Starting second pass (creation)
 >abcdefghij... |   1|  reg
                |    |    brnc
                |    |      piec
                |    |        atom
 ><             |  10|    tail~ EXACT <abcdefghijklmnopqrstuvwxyzX> (1
+) -> END
first:>  1: EXACT <abcdefghijklmnopqrstuvwxyzX> (9)
first at 1
Peep:Pos:0/0 Flags: 0x0 Whilem_c: 0 Lcp: 0 Last:'' 0:0/0 *Fixed:'' @ 0
+ Float: '' @ 0/0
Peep>  1: EXACT <abcdefghijklmnopqrstuvwxyzX> (9)
  join>  1: EXACT <abcdefghijklmnopqrstuvwxyzX> (9)
pre-fin:Pos:27/0 Flags: 0x0 Whilem_c: 0 Lcp: 0 Last:'abcdefghijklmnopq
+rstuvwxyzX' 27:0/0 *Fixed:'' @ 0 Float: '' @ 0/0
post-fin:Pos:27/0 Flags: 0x0 Whilem_c: 0 Lcp: 0 Last:'abcdefghijklmnop
+qrstuvwxyzX' 27:0/0 *Fixed:'' @ 0 Float: '' @ 0/0
commit: Pos:27/0 Flags: 0x0 Whilem_c: 0 Lcp: 0 Last:'abcdefghijklmnopq
+rstuvwxyzX' -1:0/0 *Fixed:'abcdefghijklmnopqrstuvwxyzX' @ 0 Fl
oat: '' @ 0/0
minlen: 27 r->minlen:0
Final program:
   1: EXACT <abcdefghijklmnopqrstuvwxyzX> (9)
   9: END (0)
anchored "abcdefghijklmnopqrstuvwxyzX" at 0 (checking anchored isall) 
+minlen 27
r->extflags: CHECK_ALL USE_INTUIT_NOML USE_INTUIT_ML
Guessing start of match in sv for REx "abcdefghijklmnopqrstuvwxyzX" ag
+ainst "abcdefghijklmnopqrstuvwxyzabcdefghijklmnopqrstuvwxyzabc
defgh"...
Check offset min: 0 Start shift: 0 End shift 0 Real End Shift: 0
fbm_instr len=157 str=<abcdefghijklmnopqrst>
Found anchored substr "abcdefghijklmnopqrstuvwxyzX" at offset 78...
Check offset min:0 max:0 S:78 t:78 D:0 end:157
Starting position does not contradict /^/m...
Guessed: match at offset 78
Freeing REx: "abcdefghijklmnopqrstuvwxyzX"
[download]

So, I don't understand what happens, or how the RE engine is faster than index when they both fire up fbm_instr.

52 MB and as you can see from my test I'm cosntructing a case where Boyer-Moore must be slower than brute force.

Cheers Rolf