lkml.org 
[lkml]   [2007]   [Aug]   [31]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Subject[PATCH 6/6] Use macros instead of static inline functions for zonelist iterators
Date

gcc-3.4 and probably older compiler versions produce worse code for static
inline functions than they do for macros. Due to the fact the allocator
path is a hotpath, there is approximately a 0.2% performance difference on
kernbench's System CPU times when using static inline functions. This is
not a problem on gcc 4.1.

This patch should be ignored because the static inline functions come with
type-checking and there doesn't need to be concern about macro arguements
having funky side-effects. However, as the performance problem is noticable
on older compilers, the compiler version and this patch should be checked
as the potential source and solution to a performance regression.

Signed-off-by: Mel Gorman <mel@csn.ul.ie>
---

include/linux/gfp.h | 8 ++++----
include/linux/mmzone.h | 31 +++++++++++++++++++------------
2 files changed, 23 insertions(+), 16 deletions(-)

diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.23-rc3-mm1-040_use_one_zonelist/include/linux/gfp.h linux-2.6.23-rc3-mm1-045_macro_not_inline/include/linux/gfp.h
--- linux-2.6.23-rc3-mm1-040_use_one_zonelist/include/linux/gfp.h 2007-08-31 17:22:55.000000000 +0100
+++ linux-2.6.23-rc3-mm1-045_macro_not_inline/include/linux/gfp.h 2007-08-31 17:23:10.000000000 +0100
@@ -156,11 +156,11 @@ static inline gfp_t set_migrateflags(gfp
*
* For the normal case of non-DISCONTIGMEM systems the NODE_DATA() gets
* optimized to &contig_page_data at compile-time.
+ *
+ * See the explanation above zonelist_zone() in include/linux/mmzone.h as
+ * to why this is a macro and not a static inline
*/
-static inline struct zonelist *node_zonelist(int nid)
-{
- return &NODE_DATA(nid)->node_zonelist;
-}
+#define node_zonelist(nid) (&NODE_DATA(nid)->node_zonelist)

#ifndef HAVE_ARCH_FREE_PAGE
static inline void arch_free_page(struct page *page, int order) { }
diff -rup -X /usr/src/patchset-0.6/bin//dontdiff linux-2.6.23-rc3-mm1-040_use_one_zonelist/include/linux/mmzone.h linux-2.6.23-rc3-mm1-045_macro_not_inline/include/linux/mmzone.h
--- linux-2.6.23-rc3-mm1-040_use_one_zonelist/include/linux/mmzone.h 2007-08-31 18:06:59.000000000 +0100
+++ linux-2.6.23-rc3-mm1-045_macro_not_inline/include/linux/mmzone.h 2007-08-31 17:23:10.000000000 +0100
@@ -687,15 +687,16 @@ extern struct zone *next_zone(struct zon
#endif

#define ZONELIST_ZONEIDX_MASK ((1UL << ZONES_SHIFT) - 1)
-static inline struct zone *zonelist_zone(unsigned long zone_addr)
-{
- return (struct zone *)(zone_addr & ~ZONELIST_ZONEIDX_MASK);
-}

-static inline int zonelist_zone_idx(unsigned long zone_addr)
-{
- return zone_addr & ZONELIST_ZONEIDX_MASK;
-}
+/*
+ * Subtle: These are macros, not static inlines because gcc 3.4 at least
+ * produces worse code with static inline functions. The effect is about 0.4%
+ * regression in kernbench tests. The problem doesn't appear to exist on
+ * gcc 4.1
+ */
+#define zonelist_zone_idx(zone_addr) ((zone_addr) & ZONELIST_ZONEIDX_MASK)
+#define zonelist_zone(zone_addr) \
+ ((struct zone *)((zone_addr) & ~ZONELIST_ZONEIDX_MASK))

static inline unsigned long encode_zone_idx(struct zone *zone)
{
@@ -706,13 +707,19 @@ static inline unsigned long encode_zone_
return encoded;
}

-static inline int zone_in_nodemask(unsigned long zone_addr, nodemask_t *nodes)
-{
#ifdef CONFIG_NUMA
- return node_isset(zonelist_zone(zone_addr)->node, *nodes);
+#define zone_in_nodemask(zone_addr, nodes) \
+ (node_isset(zonelist_zone(zone_addr)->node, *nodes))
#else
- return 1;
+#define zone_in_nodemask(zone_addr, nodes) (1)
#endif /* CONFIG_NUMA */
+
+static inline int zone_in_nodemask(unsigned long zone_addr, nodemask_t *nodes)
+{
+ if (NUMA_BUILD)
+ return node_isset(zonelist_zone(zone_addr)->node, *nodes);
+
+ return 1;
}

/* Returns the first zone at or below highest_zoneidx in a zonelist */
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
\
 
 \ /
  Last update: 2007-08-31 22:57    [W:0.068 / U:0.932 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site