[Geany-Devel] Implementing stdlib/glib in Geany

[Colomban Wendling]

Le 27/04/2014 22:28, Pavel Roschin a écrit :
> I found interesting function in utils.c:
> 
> gboolean utils_str_equal(const gchar *a, const gchar *b)
> {
> 	[...]
> 
> 	while (*a == *b++)
> 		if (*a++ == '\0')
> 			return TRUE;
> 
> 	return FALSE;
> }
> 
> This function is widely used in Geany code.
> 
> However:
> - it is not inline so there is no reason to keep this small while loop
> - it uses own while-loop to compare that is *slow* for big strings
> - GLib has similar g_strcmp0 function since 2.16
> 
> Using while-loop may confuse compiler which knows strcmp and may optimize this
> std call e.g. for constants. Furthermore strcmp is SSE-accelerated function
> which compares 128 bits per cycle while this implementstion may compare only 8
> bits per cycle.

Ouch indeed.  I just replaced the loop with a strcmp(), which performs
roughly 5.4 times better (on a 64bits Linux) for dummy average strings.
Thanks for spotting this.

> Replace this by inline (g_strcmp0(a, b) == 0). This is much faster and easier to read.

Well, the problem is that utils_str_equal() is part of the plugin API,
so we can't make it a macro expanding to a non-function without breaking
API.
Also, I don't think "g_strcmp0(a, b) == 0" is easier to read than
"utils_str_equal(a, b)" -- OK strcmp() is obvious, but strcmp0() is not,
and utils_str_equal() is quite explicit.

Also, my naive tests showing the strcmp() version being 5.4 times faster
(from 22.8301s to 4.22474s) shows very little gain with calling
g_strcmp0() (+.02, 42.1628s vs 41.242s) -- and I would think it probably
depends on optimization/debug compiler options, since both
implementations are very similar (maybe g_strcmp0() is a little better
because it does some tests only in some branches, but OTOH the compiler
can easily improve our branching to lower test hitting).

Regards,
Colomban