Re: PR 51386

[François Dumont <frs dot dumont at gmail dot com>]

On 12/05/2011 10:13 PM, Paolo Carlini wrote:
> Hi,
>
> >     Tested under linux x86_64, load_factor test is much faster now.
>
> How much, exactly?

About 2000 times faster. I have temporarily added the slow part of 
load_factor test in a performance test to measure it, do you want to 
keep this ? I hadn't proposed it before because a max load factor of 0.3 
is quite unusual. load_factor test simply challenge a possible use case 
that has little chances to be experimented in a production environment.

> >
> > 2011-12-05  François Dumont <fdumont@gcc.gnu.org>
> >
> >         PR 51386
>
> PR libstdc++/51386
>
> >
> >         * include/bits/hashtable_policy.h 
> (_Prime_rehash_policy::_M_next_bkt):
> >         Fix computation of _M_prev_resize so that hashtable do not 
> keep on
> >         being rehashed when _M_max_load_factor is lower than 1.
> >
> > Ok to commit ?
>
> The whole patch doesn't really look self-explaining, but I'm willing 
> to largely trust you ;) Please improve a bit the English in the comments:
>
> > Index: include/bits/hashtable_policy.h
> > ===================================================================
> > --- include/bits/hashtable_policy.h    (revision 181975)
> > +++ include/bits/hashtable_policy.h    (working copy)
> > @@ -298,25 +298,34 @@
> >    _Prime_rehash_policy::
> >    _M_next_bkt(std::size_t __n) const
> >    {
> > -    // Optimize lookups involving the first elements of __prime_list.
> > -    // (useful to speed-up, eg, constructors)
> > -    static const unsigned long __fast_bkt[12]
> > -      = { 2, 2, 2, 3, 5, 5, 7, 7, 11, 11, 11, 11 };
> > +    // Optimize lookups involving the first elements of 
> __prime_list (useful to
> > +    // speed-up, eg, constructors). The contained values are 
> already containing
> > +    // a minimal grow step.
>
> I would say "The values already include a minimal grow step" and avoid 
> the repetition. But to be honest, I don't like much that the numbers 
> are not the same numbers as in __prime_list. This should be *purely* a 
> trick to improve the speed of the search, should not serve other 
> purposes. Can we decouple the two things? Sooner or later the 
> inconsistency would byte us, I'm sure. Please reason this way: imagine 
> the trick was not there, re-do the patch, and, at the *very* end, put 
> the trick back *only* to avoid calling lower_bound.
I finally restore the original values, the small containers will simply 
grow slower at the beginning, not a big deal.

Ok to commit ?

François

Index: include/bits/hashtable_policy.h
===================================================================
--- include/bits/hashtable_policy.h	(revision 181975)
+++ include/bits/hashtable_policy.h	(working copy)
@@ -300,23 +300,30 @@
   {
     // Optimize lookups involving the first elements of __prime_list.
     // (useful to speed-up, eg, constructors)
-    static const unsigned long __fast_bkt[12]
+    static const unsigned char __fast_bkt[12]
       = { 2, 2, 2, 3, 5, 5, 7, 7, 11, 11, 11, 11 };
 
+    if (__n <= 11)
+      {
+	_M_prev_resize = 0;
+	_M_next_resize
+	  = __builtin_ceil(__fast_bkt[__n] * (long double)_M_max_load_factor);
+	return __fast_bkt[__n];
+      }
+
     const unsigned long* __p
-      = __n <= 11 ? __fast_bkt + __n
-		  : std::lower_bound(__prime_list + 5,
-				     __prime_list + _S_n_primes, __n);
+      = std::lower_bound(__prime_list + 5, __prime_list + _S_n_primes, __n);
 
-    _M_prev_resize = __builtin_floor(*__p * (long double)_M_max_load_factor);
-    if (__p != __fast_bkt)
-      _M_prev_resize = std::min(_M_prev_resize,
-				static_cast<std::size_t>(*(__p - 1)));
-    // Lets guaranty a minimal grow step of 11:
+    // Shrink will take place only if the number of elements is small enough
+    // so that the prime number 2 steps before __p is large enough to still
+    // conform to the max load factor:
+    _M_prev_resize
+      = __builtin_floor(*(__p - 2) * (long double)_M_max_load_factor);
+
+    // Let's guaranty that a minimal grow step of 11 is used
     if (*__p - __n < 11)
-      __p = std::lower_bound(__prime_list + 5,
-			     __prime_list + _S_n_primes, __n + 11);
-    _M_next_resize = __builtin_floor(*__p * (long double)_M_max_load_factor);
+      __p = std::lower_bound(__p, __prime_list + _S_n_primes, __n + 11);
+    _M_next_resize = __builtin_ceil(*__p * (long double)_M_max_load_factor);
     return *__p;
   }