Re: [RFC/PATCH 4/5] completion: get rid of compgen

On Sat, Nov 17, 2012 at 3:12 PM, SZEDER Gábor [off-list ref] wrote:

On Sat, Nov 17, 2012 at 12:42:38PM +0100, Felipe Contreras wrote:

quoted

On Sat, Nov 17, 2012 at 12:00 PM, SZEDER Gábor [off-list ref] wrote:

quoted

On Sat, Nov 17, 2012 at 02:38:17AM +0100, Felipe Contreras wrote:

quoted

The functionality we use is very simple, plus, this fixes a known
breakage 'complete tree filename with metacharacters'.

Signed-off-by: Felipe Contreras <redacted>
---
 contrib/completion/git-completion.bash | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/contrib/completion/git-completion.bash b/contrib/completion/git-completion.bash
index 975ae13..ad3e1fe 100644
--- a/contrib/completion/git-completion.bash
+++ b/contrib/completion/git-completion.bash

@@ -227,7 +227,11 @@ fi

 __gitcompadd ()
 {
-     COMPREPLY=($(compgen -W "$1" -P "$2" -S "$4" -- "$3"))
+     for x in $1; do
+             if [[ "$x" = "$3"* ]]; then
+                     COMPREPLY+=("$2$x$4")
+             fi
+     done

The whole point of creating __gitcomp_nl() back then was to fill
COMPREPLY without iterating through all words in the wordlist, making
completion faster for large number of words, e.g. a lot of refs, or
later a lot of symbols for 'git grep' in a larger project.

The loop here kills that optimization.

So your solution is to move the loop to awk? I fail to see how that
could bring more optimization, specially since it includes an extra
fork now.

This patch didn't aim for more optimization, but it was definitely a
goal not to waste what we gained by creating __gitcomp_nl() in
a31e6262 (completion: optimize refs completion, 2011-10-15).  However,
as it turns out the new version with awk is actually faster than
current master with compgen:

  Before:

    $ refs="$(for i in {0..9999} ; do echo branch$i ; done)"
    $ time __gitcomp_nl "$refs"

    real    0m0.242s
    user    0m0.220s
    sys     0m0.028s

  After:

    $ time __gitcomp_nl "$refs"

    real    0m0.109s
    user    0m0.096s
    sys     0m0.012s

This one is even faster:

while read -r x; do
	if [[ "$x" == "$3"* ]]; then
		COMPREPLY+=("$2$x$4")
	fi
done <<< $1

== 10000 ==
one:
real	0m0.148s
user	0m0.134s
sys	0m0.025s
two:
real	0m0.055s
user	0m0.054s
sys	0m0.000s

-- 
Felipe Contreras