Re: Omitted conversion float -> double in x86 back-end

[Michael Matz <matz at suse dot de>]

Hi,

On Tue, 21 Nov 2006, Richard Guenther wrote:

> but then regstack comes along and
> 
> (insn:TI 10 32 33 2 (set (reg:SF 8 st)
>        (float:SF (mem/c/i:DI (plus:SI (reg/f:SI 6 bp)
>                    (const_int 8 [0x8])) [3 l+0 S8 A32]))) 172
> {*floatdisf2_i387} (nil)
>    (nil))
> 
> where the *floatdisf2_i387 pattern misses the truncation to SFmode.
> 
> (you should file a bugzilla about this)

No you don't need to.  This is the usual x87-does-everything-in-extended 
problem, i.e. PR 323.  There is no way to truncate that extended precision 
to float (or double for that matter) without explicitely storing to 
memory.  If you want to take that performance hit use -ffloat-store, in 
which case the correct code will be produced:

foo:
        pushl   %ebp
        movl    %esp, %ebp
        subl    $16, %esp
        fildll  8(%ebp)
        fstps   -4(%ebp)
        flds    -4(%ebp)
        fstpl   -16(%ebp)
        fldl    -16(%ebp)
        leave
        ret

Better of course if to use SSE math.


Ciao,
Michael.