| commit | eb2cd968ffbb137a6c4bf8c37917972536559a4a | |
| author | Rodrigo Kumpera <kumpera@gmail.com> | |
| Thu, 17 Jun 2010 19:24:01 +0000 (17 19:24 -0000) | ||
| committer | Rodrigo Kumpera <kumpera@gmail.com> | |
| Thu, 17 Jun 2010 19:24:01 +0000 (17 19:24 -0000) | ||
| tree | da3b2211c941195e88d9d1c39327daa4cac22d8e | treesnapshot (tar.gz zip) |
| parent | 81dfccbd14f683f1fe46f570e76c1e99a14c5a81 | commitdiff |
2010-06-17 Rodrigo Kumpera <rkumpera@novell.com>
* method-to-ir.c (mini_emit_stobj): Don't call mono_value_copy for small
value types as the overhead is huge. Manually expand it up to 5 words to
avoid code bloat.
This improves gmcs times by 5% and unmodified binary-tree by 78%. The later
is an exception that performance is dominated by mono_value_copy.
This puts sgen about 5% ahead of boehm in terms of wall-clock on a Core2Quad.
svn path=/trunk/mono/; revision=159081
* method-to-ir.c (mini_emit_stobj): Don't call mono_value_copy for small
value types as the overhead is huge. Manually expand it up to 5 words to
avoid code bloat.
This improves gmcs times by 5% and unmodified binary-tree by 78%. The later
is an exception that performance is dominated by mono_value_copy.
This puts sgen about 5% ahead of boehm in terms of wall-clock on a Core2Quad.
svn path=/trunk/mono/; revision=159081
| mono/mini/ChangeLog | diffblobblamehistory | |
| mono/mini/method-to-ir.c | diffblobblamehistory |