Andy Polyakov <[EMAIL PROTECTED]> wrote:
> I was just about to suggest to rename solaris-sparc-cc to
> solaris-sparc-sc3 and solaris-*-sc4 to solaris-*-cc which would cover
> both SC4.x and SC5.x... But I can see you've introduced sc5 option:-(

This happened last week.  Someone noticed that SC4.2 (with all the latest
patches) with -xtarget=ultra -xO5 generates bad code.  Exptest breaks.
SC5.0 (with the compiler patches) works.  And removing -xtarget=ultra or
turning it down to -xO2 makes 4.2 work.  Hence the distinction between SC4
and SC5.

Which should be the "default" SC version, 4 or 5?  I like 5, which means
we would change it so
    SC3 = solaris-*-sc3
    SC4 = solaris-*-sc4
    SC5 = solaris-*-cc
Or if most people will stick with SC4 for a while longer then we could make
it -sc3, -cc, and -sc5.

> In either case I don't find my v8plus implementation of bn_asm.c in
> solaris-usparc-* lines. How come? After all it performs at least 25%
> faster than asm/sparcv8.o on UltraSPARC. Also note that SC5.0 in turn
> generates faster code than asm/sparcv8.o. I.e. if for any reason you
> don't want to integrate v8plus module, then at least asm/sparcv8.o
> should be taken away from solaris-usparc-sc5 line.

The current -sc4/-sc5 configurations are what you suggested to the list.
At least that was the intent.  Your v8plus asm isn't in there simply because
the .S file never made it into CVS.  Please repost the asm file and diffs
to make the configuration lines the way they ought to be.

> SC4/5 provides another cool flag worth testing that makes some things
> run faster, namely -xrestrict. The flags tells compiler to relax and
> don't assume any data aliasing so that it reschedules loads/stores in
> more daring way. Well, it does mean that programmer has to take special
> care about cases where aliasing is intentional. In either case I've
> observed 20% performance improvement in aligned data case in MD5 I'm
> working on right now. Well, of course with SC5.0...

That seems too risky to do it project-wide.  Do we want to go back to the
old way - have asm files that are just output from the fastest compiler?

______________________________________________________________________
OpenSSL Project                                 http://www.openssl.org
Development Mailing List                       [EMAIL PROTECTED]
Automated List Manager                           [EMAIL PROTECTED]

Reply via email to