Kevin D. Kissell wrote:
> > Nigel Stephens wrote:
> > > Thiemo Seufer wrote:
> > > >GNU as has "move" as builtin macro which is expanded differently for
> > > >32 and 64 bit mode. This simplifies the task of writing code for both
> > > >models. Unfortunately the expansion was broken for a while and
> > > >generated always the 64 bit version if the toolchain was 64bit
> > > >_capable_. IIRC this was fixed in the early 2.14 timeframe.
> > >
> > > Wouldn't it be safer, as Kevin suggests, for the "move" macro to
> > > generate "or $rd,$0,$rs", since that will work correctly in either mode?
> >
> > For CPUs with two adders the instruction scheduling is a bit better.
>
> For CPUs with two adders, only one of which can do ORs, you mean.
Yes.
> Isn't that a pretty small population of MIPS CPUs?
Sure, but there's little reason to deny this optimization.
> > Ther are also many other macros which are expanded in dependence of
> > the 32/64 bit address/register size, a different "move" macro
> > expansion won't help much if the assembler invocation was wrong.
>
> Perhaps, but it strikes me as being only prudent to remove unnecessarily
> fragile artifacts like this.
It was an assembler bug.
Thiemo
|