Mailing List Archive

default fp selection with armv6/armv7 targets
currently, we have the system:
- if chost matches *-softfloat-*, you get softfloat
- if chost matches *-hardfloat-*, you get hardfloat
- for everything else, you get the gcc default

with the standardization work going on with armv7+ and hardfp, i've made the
following change:
- if chost matches armv7*-softfloat-*, you got softfloat
- for all other armv7* targets, you get hardfloat vfp3-d16

along those lines, i've also slipped in:
- if chost matches armv6*-softfloat-*, you got softfloat
- for all other armv6* targets, you get hardfloat vfp

considering vfp is required baseline in these cores now, it doesn't make sense
to not use it if the user has explicitly stated they're targeting these
arches.

if you really want to use a different default, you can still use EXTRA_ECONF to
set whatever you want.
-mike
Re: default fp selection with armv6/armv7 targets [ In reply to ]
Using the least common denominator is another food reason to consider link
time optimization... I wonder if e.g. fatelf would make it easier to
implement that. Multiple arm binaries glued together based on variant and
hwcaps (e.g. v6+vfp-d16, v7+vfp-d32, v7+neon)? Heck, if fatelf were used,
only one stage3 tar ball would be necessary for all arch's combined...
although downloading would be significantly slower :-/

©

Sent from my Android
On May 18, 2012 1:10 AM, "Mike Frysinger" <vapier@gentoo.org> wrote:

> currently, we have the system:
> - if chost matches *-softfloat-*, you get softfloat
> - if chost matches *-hardfloat-*, you get hardfloat
> - for everything else, you get the gcc default
>
> with the standardization work going on with armv7+ and hardfp, i've made
> the
> following change:
> - if chost matches armv7*-softfloat-*, you got softfloat
> - for all other armv7* targets, you get hardfloat vfp3-d16
>
> along those lines, i've also slipped in:
> - if chost matches armv6*-softfloat-*, you got softfloat
> - for all other armv6* targets, you get hardfloat vfp
>
> considering vfp is required baseline in these cores now, it doesn't make
> sense
> to not use it if the user has explicitly stated they're targeting these
> arches.
>
> if you really want to use a different default, you can still use
> EXTRA_ECONF to
> set whatever you want.
> -mike
>