[Bug target/79593] [6/7 Regression] Poor/Worse code generation for FPU on versions after 6

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79593

--- Comment #9 from rguenther at suse dot de <rguenther at suse dot de> ---
On Tue, 21 Feb 2017, ubizjak at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79593
> 
> Uroš Bizjak <ubizjak at gmail dot com> changed:
> 
>            What    |Removed                     |Added
> ----------------------------------------------------------------------------
>                  CC|                            |rguenth at gcc dot gnu.org
> 
> --- Comment #8 from Uroš Bizjak <ubizjak at gmail dot com> ---
> The remaining fld %st(0) is there due to multiple uses of global_data+0, once
> in XFmode and once in SFmode. But this is what tree optimizers give us:
> 
>   long double min;
>   long double delta;
> 
>   float _2;
> 
>   <bb 3> [54.00%]:
>   _2 = global_data[0];
>   delta_13 = (long double) _2;
>   _3 = global_data[1];
>   min_14 = (long double) _3;
>   pretmp_28 = *e_11(D).D.2291.f;
>   if (_2 < 0.0)
>     goto <bb 4>; [36.00%]
>   else
>     goto <bb 5>; [64.00%]
> 
> Please note the comparison (_2 < 0.0). I don't know why, looking at the source,
> delta_13 should be used there. Let's ask tree experts.

Simply an optimization (already the FEs via convert_to_real narrow a
widened comparison).