lib: bitmap: make nbits parameter of bitmap_empty unsigned
authorRasmus Villemoes <linux@rasmusvillemoes.dk>
Wed, 6 Aug 2014 23:09:49 +0000 (16:09 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 7 Aug 2014 01:01:25 +0000 (18:01 -0700)
commit0679cc483669d08153d158273455398a389ee9ca
tree1212e8b313cda2a96141c889436d6680f72befa1
parentd0da23b0debcef135c866cc8117d197fb40a6079
lib: bitmap: make nbits parameter of bitmap_empty unsigned

Many functions in lib/bitmap.c start with an expression such as lim =
bits/BITS_PER_LONG.  Since bits has type (signed) int, and since gcc
cannot know that it is in fact non-negative, it generates worse code
than it could.  These patches, mostly consisting of changing various
parameters to unsigned, gives a slight overall code reduction:

  add/remove: 1/1 grow/shrink: 8/16 up/down: 251/-414 (-163)
  function                                     old     new   delta
  tick_device_uses_broadcast                   335     425     +90
  __irq_alloc_descs                            498     554     +56
  __bitmap_andnot                               73     115     +42
  __bitmap_and                                  70     101     +31
  bitmap_weight                                  -      11     +11
  copy_hugetlb_page_range                      752     762     +10
  follow_hugetlb_page                          846     854      +8
  hugetlb_init                                1415    1417      +2
  hugetlb_nrpages_setup                        130     131      +1
  hugetlb_add_hstate                           377     376      -1
  bitmap_allocate_region                        82      80      -2
  select_task_rq_fair                         2202    2191     -11
  hweight_long                                  66      55     -11
  __reg_op                                     230     219     -11
  dm_stats_message                            2849    2833     -16
  bitmap_parselist                              92      74     -18
  __bitmap_weight                              115      97     -18
  __bitmap_subset                              153     129     -24
  __bitmap_full                                128     104     -24
  __bitmap_empty                               120      96     -24
  bitmap_set                                   179     149     -30
  bitmap_clear                                 185     155     -30
  __bitmap_equal                               136     105     -31
  __bitmap_intersects                          148     108     -40
  __bitmap_complement                          109      67     -42
  tick_device_setup_broadcast_func.isra         81       -     -81

[The increases in __bitmap_and{,not} are due to bug fixes 17/18,18/18.
No idea why bitmap_weight suddenly appears.] While 163 bytes treewide is
insignificant, I believe the bitmap functions are often called with
locks held, so saving even a few cycles might be worth it.

While making these changes, I found a few other things that might be
worth including.  16,17,18 are actual bug fixes.  The rest shouldn't
change the behaviour of any of the functions, provided no-one passed
negative nbits values.  If something should come up, it should be fairly
bisectable.

A few issues I thought about, but didn't know what to do with:

* Many of the functions misbehave if nbits is compile-time 0; the
  out-of-line functions generally handle 0 correctly.  bitmap_fill() is
  particularly bad, whether the 0 is known at compile time or not.  It
  would probably be nice to add detection of at least compile-time 0 and
  handle that appropriately.

* I didn't change __bitmap_shift_{left,right} to use unsigned because I
  want to fully understand why the algorithm works before making that
  change.  However, AFAICT, they behave correctly for all (positive) shift
  amounts.  This is not the case for the small_const_nbits versions.  If
  for example nbits = n = BITS_PER_LONG, the shift operators turn into
  no-ops (at least on x86), so one get *dst = *src, whereas one would
  expect to get *dst=0.  That difference in behaviour is somewhat
  annoying.

This patch (of 18):

The compiler can generate slightly smaller and simpler code when it
knows that "nbits" is non-negative.  Since no-one passes a negative
bit-count, this shouldn't affect the semantics.

Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/linux/bitmap.h
lib/bitmap.c