Re: [Ikvm-developers] Float problems

[Dennis U. <ls...@ya...>]

Hi.

You wrote:
JF> Please be careful. This license on that file doesn't allow modification.
Oh... Didn't notice. Actually it has an error at line 821, it's float
wide cancellation test. They sum 2 values and then subtract another
one without storing intermediate result in local variable. This is
correct only for strict mode. In non-strict, sum must be stored in a
local variable, so it can be truncated. It passes on JVM because all
known JVM's are working in strict mode.

JF> Interesting. Where are you doing the conv.r4/8? I would expect
JF> them to be necessary only after a mul/div/rem/add/sub operation.
Adding them after arithmetic operations is correct, but it gives too
much performance degrade (see ikvm_scimark_float_fix_full_strict.log)
and it's only required for strict mode. So, I add conv.r4/8 only after
'escape'-operations like dstore/fstore, dup, putfield/putstatic except
dastore/daload (because it has no impact on my tests - FloatTest and
DaCapo.lusearch and it's too slow). This is compliant with Java
Specification.

Also I've improved compare methods. They no longer need that pack of
branch instructions. Result is based on that fact:
cmp(a,b) = cgt(a,b) - clt(a,b)

JF> I'm not a big fan of command line switches. So my initial thought
JF> would be to only do "compliant", but I haven't made up my mind yet.
If you decide that switches are necessary, write me, I can add this
functionality.


-- 
Bye,
 Dennis                          mailto:ls...@ya...