RE: [patch] Make fixed-point multiplication results consistent

["Fu, Chao-Ying" <fu at mips dot com>]

Mark Mitchell wrote:
> 
> >   Currently constant folding for fixed-point multiplications
> > is not consistent with MIPS instructions (ex: "mulq_rs.w")
> > due to rounding styles.
> 
> What does the draft ISO spec for fixed-point say about this case?  Is
> there a standard rounding mode, or is this like floating-point where
> there are a variety of modes?

  From the spec (n1169.pdf):
Page 12:
"If the source value cannot be represented exactly by the fixed-point type,
the source value is rounded to either the closest fixed-point value greater
than the source value (rounded up) or to the closest fixed-point value less 
than the source value (rounded down)."

Page 13:
"Whether rounding is up or down is implementation-defined and may differ 
for different values and different situations; an implementation may 
specify that the rounding is indeterminable."

  So, each target can have its own rounding mode or even let the rounding
indeterminable.

> 
> The reason that I'm asking is that the goal of the fixed-point code
> should not be to match the MIPS processor -- it should be to match the
> spec.  At least in principle, that might not be the same 
> thing in every
> case.  And, if the spec allows multiple modes, then probably the
> emulation code should do so as well.
> 
> I'm trying to figure out whether your change is correct as it 
> stands, or
> if it needs to be conditionalized in some way.
> 

  We may use a target hook/macro to conditionally perform one specific
rounding mode for one target in constant folding and emulation.
Thanks!

Regards,
Chao-ying