Re: [PATCH, midlevel]: Handle (int)trunc(float x) cases

[Uros Bizjak <uros at kss-loka dot si>]

Roger Sayle wrote:

> Sorry for the delay in reviewing but I've been investigating the
> asymmetry between trunc and the other integer rounding functions;
> round, floor and ceil.  The following code from expr.c explains a
> lot:
>
>  
>
> > >   case FIX_ROUND_EXPR:
> > >   case FIX_FLOOR_EXPR:
> > >   case FIX_CEIL_EXPR:
> > >     gcc_unreachable ();                       /* Not used for C.  */
> > >
> > >   case FIX_TRUNC_EXPR:
> > >     op0 = expand_expr (TREE_OPERAND (exp, 0), NULL_RTX, VOIDmode, 0);
> > >     if (target == 0 || modifier == EXPAND_STACK_PARM)
> > >       target = gen_reg_rtx (mode);
> > >     expand_fix (target, op0, unsignedp);
> > >     return target;
> > >      
>
> So although tree codes are defined for the other "fix" conversions,
> it's clearly unsafe to use them.  It's not clear what the best way
> to clean this up is (either add to add the missing support or to
> remove the partially implemented tree codes).  An issue for another
> day...
>
>  
I have been trying to implement these conversions in the same way as 
FIX_TRUNC_EXPR (looking at the same code as you quote above). But to 
implement these conversion as fix_trunc[INTmode][FPmode] optab, a lot of 
expanders would be needed.

As a compromise, the same functionality could be acheived by 
implementing only a  fix_floor[INTmode] expanders, where its FPmode 
input is always in widest FP mode (XFmode in i386 case). Please note 
that this expansion is performed in expand_unop. As lfloor and ltrunc 
optabs are enabled only for -ffast-math, where FPmode->XFmode conversion 
is 'free', this results in quite optimal asm code.

Uros.