[Gmp-commit] /var/hg/gmp: Fix timing table and add some comments.

[mercurial at gmplib.org]

details:   /var/hg/gmp/rev/1c891bd40cff
changeset: 14021:1c891bd40cff
user:      Torbjorn Granlund <tege at gmplib.org>
date:      Wed Mar 09 22:11:17 2011 +0100
description:
Fix timing table and add some comments.

diffstat:

 mpn/x86_64/aorrlsh_n.asm |  11 +++++++----
 1 files changed, 7 insertions(+), 4 deletions(-)

diffs (24 lines):

diff -r 22d94377cc74 -r 1c891bd40cff mpn/x86_64/aorrlsh_n.asm
--- a/mpn/x86_64/aorrlsh_n.asm	Wed Mar 09 21:37:26 2011 +0100
+++ b/mpn/x86_64/aorrlsh_n.asm	Wed Mar 09 22:11:17 2011 +0100
@@ -22,13 +22,16 @@
 C	     cycles/limb
 C AMD K8,K9	 2.87	< 3.85 for lshift + add_n
 C AMD K10	 2.75	< 3.85 for lshift + add_n
-C Intel P4	14.6	> 7.33 for lshift + add_n
-C Intel core2	 4	> 3.27 for lshift + add_n
-C Intel NHM	 2.83	< 3.75 for lshift + add_n
-C Intel SBR	 3.25	< 3.46 for lshift + add_n
+C Intel P4	22	> 7.33 for lshift + add_n
+C Intel core2	 4.1	> 3.27 for lshift + add_n
+C Intel NHM	 4.4	< 3.75 for lshift + add_n
+C Intel SBR	 3.17	< 3.46 for lshift + add_n
 C Intel atom	 ?	< 8.75 for lshift + add_n
 C VIA nano	 4.7	< 6.25 for lshift + add_n
 
+C TODO
+C  * Can we propagate carry into rdx instead of using a special carry register?
+C    That could save enough insns to get to 10 cycles/iteration.
 
 define(`rp',       `%rdi')
 define(`up',       `%rsi')