🚀 go-pugleaf

In article <3135@jade.BERKELEY.EDU>, desj@lemon.BERKELEY.EDU (David desJardins) writes:
> In article <3300003@uiucdcsm> grunwald@uiucdcsm.cs.uiuc.edu writes:
> >
> >Strength reduction is replacement of operations such as "i = j * 8" by
> >"i = j << 3", or "i = j % 8" by "i = j & 0x3" (if j is unsigned).
>                                   ^^^^^^^^^^^   ^^^^^^^^^^^^^^^^
>    I'm sure you mean "i = j & 0x7".  But, more importantly, this is the
> right remainder to compute even if j is signed (assuming two's-complement,
> of course).  It's too bad more people (especially hardware designers!)
> don't realize this.

I think you should be more explicit in what you are stating here.  The
strength-reduction optimization above does *not* work for signed dividends
(j, in this case) on most current machines, since they define the remainder
to have the same sign as the dividend.  Thus, you seem to be arguing that
this definition be changed to one where remainders are always positive and
integer divide operations always round *down* instead of towards zero (this
is required for 

	(a/b)*b + a%b	==	a


I seem to remember reading that the PL.8 compiler for the IBM ROMP chip
used this definition for divide operations so that the compiler could
perform this very optimization.

There was a discussion a while (1 yr?) ago about this subject, but I can't
remember what conclusions were made (if any).  Anyone care to refresh our
memories?

	Tim Olson
	Advanced Micro Devices
	(tim@amdcad.AMD.COM)