Re: [PATCH] PR target/65240, Fix Power{7,8} insn constraint issue with -O3 -ffast-math

[David Edelsohn <dje dot gcc at gmail dot com>]

On Thu, Mar 12, 2015 at 11:29 AM, Michael Meissner
<meissner@linux.vnet.ibm.com> wrote:
> On Wed, Mar 11, 2015 at 08:52:54PM -0400, David Edelsohn wrote:
>> On Wed, Mar 11, 2015 at 6:21 PM, Michael Meissner
>> <meissner@linux.vnet.ibm.com> wrote:
>> > On Wed, Mar 11, 2015 at 01:02:06PM -0400, David Edelsohn wrote:
>> >> I am concerned with the create_TOC_reference use for TARGET_TOC.  Has
>> >> this been tested with big endian -mcmodel=small?
>> >
>> > Yes, that was a problem.  Patch coming up soon.  Thanks.
>>
>> Can you call rs6000_emit_move_directly?
>
> Well, I can, but I would have to have some sort of flag that says after the
> split1 pass not to allow FP constants in move (other than 0.0).  It is doable,
> but it does touch more areas in the rs6000 back end.
>
> I am starting to think that it is just simpler to rip out all of the special
> fast math handling of constants, considering the multiply by reciprocal support
> has moved to SSA/tree and away from RTL.  Did you want me to investigate the
> performance implications of removing it now (rather than waiting to GCC 6.0),
> or just do the more limited patch that I've been pursuing.

Please check on the performance implications of removing the special
constant support.  I know that it is late, but I think that ripping it
out is less risky than trying to fix this, if the performance impact
is not bad.

Thanks, David