325657 - Unset ECL_USE_FP for USE_ABI32_INT64 and USE_ABI64_INT Solaris SPARC freebl libraries

Status: RESOLVED FIXED

(NSS :: Libraries, defect, P1)

Description

[Wan-Teh Chang]

In mozilla/security/nss/lib/freebl/Makefile, we set
ECL_USE_FP=1 for all Solaris SPARC freebl libraries.
Since some UltraSPARC processors do not have strong
floating point performance, we should not set
ECL_USE_FP when we build the freebl libraries
(USE_ABI32_INT64 and USE_ABI64_INT) for these
processors.

Nelson, please assign this bug to the appropriate
Sun NSS developer because I don't have any systems
that use these UltraSPARC processors.  The patch to
implement this change is trivial, but we need to
measure the performance to be sure this change is
a win.

Comment 1

[Nelson Bolyard (seldom reads bugmail)]

Neil, please give this top priority.  

Comment 2

[Neil Williams]

Attachment: proposed patch

I think patch does what was suggested but I need a decent performance test for ECC. Any suggestions?

Comment 3

[Wan-Teh Chang]

It's not clear to me whether we should set ECL_USE_FP=1 or not
for the USE_ABI32_INT32 freebl library.

Comment 4

[Nelson Bolyard (seldom reads bugmail)]

Comment on attachment 210670 [details] [diff] [review]
proposed patch

This patch adds 4 lines, 
two in the part that is GCC specific, and 
two in the part that is SunStudio specific, and 
none in the part that is common to GCC and Studio.  
I believe the two lines can be added into the common part, 
and are not needed in the two compiler-specific parts.

Wan-Teh wondered about USE_ABI32_INT32.  We haven't used FPU 
for that process/ABI before, so for now, let's continue to not
use the FPU on it.

Comment 5

[Neil Williams]

Attachment: incorporates Nelson's suggestion

Bltest has an option to test ECC signing. I'll post another comment as soon as I have some before-and-after-this-patch results.

Comment 6

[Nelson Bolyard (seldom reads bugmail)]

Comment on attachment 211031 [details] [diff] [review]
incorporates Nelson's suggestion

r=nelson
Floating point performance on certain CPUs is SO BAD that I'm confident this change will improve things.  
be sure that any comparative measurements are done on those CPUs.

Comment 7

[Wan-Teh Chang]

Comment on attachment 211031 [details] [diff] [review]
incorporates Nelson's suggestion

r=wtc.

Comment 8

[Neil Williams]

Checking in Makefile;
/cvsroot/mozilla/security/nss/lib/freebl/Makefile,v  <--  Makefile
new revision: 1.72; previous revision: 1.71

Comment 9

[Wan-Teh Chang]

Neil, this patch also needs to go into the NSS_3_11_BRANCH.

Comment 10

[Neil Williams]

3.11 branch
Checking in Makefile;
/cvsroot/mozilla/security/nss/lib/freebl/Makefile,v  <--  Makefile
new revision: 1.70.2.2; previous revision: 1.70.2.1
done

Thanks, Wan-Teh