Purra Joel (TCS) 2011-01-10 15:37:23 EET

STEPS LEADING TO PROBLEM:
--------------------------
floating point compiler optimization has been enabled but the effect was less
than expected

It is caused by the function int qRound(qreal d) , defined in qglobal.h

Reason is as follows :
------------------------
Code like below line cause conversion to double
qreal b = 3.0;
qreal a = 2.0* b;
double calculations are not optimized

It should be
qreal b = 3.0f;
qreal a = 2.0f* b;

For more on double conversion analysis, refer
https://projects.maemo.org/mediawiki/index.php/SystemAnalysis/DoubleConversion

Comment 1 Liimatainen Jari Juhani (Nokia) error_managers 2011-01-10 15:50:45 EET

Joel sent me the actual code of qRound via e-mail:

inline int qRound(qreal d) // here, qreal can be though of as float for ARM.

{ return d >= 0.0 ? int(d + 0.5) : int(d - int(d-1) + 0.5) + int(d-1); }

If this is called all over our code, it would be nice to get this fixed indeed,
as it causes unnecessary double conversions, and results into not using
run-fast mode (i.e. NEON), as it does not support doubles, but only floats.

Comment 2 Mannisto Tatu (Nokia) 2011-01-10 16:07:58 EET

You may also look at
https://projects.maemo.org/mediawiki/index.php?title=SystemAnalysis/NEON%2Bfloats
for other potential ways to handle these small Qt math related functions.

As stated there the 0.5 constant should be used as qreal(0.5) in the qRound.

Comment 3 Pakarinen Juuso (Tieto) 2011-01-13 09:40:24 EET

(In reply to comment #2)
> You may also look at
> https://projects.maemo.org/mediawiki/index.php?title=SystemAnalysis/NEON%2Bfloats
> for other potential ways to handle these small Qt math related functions.
>
> As stated there the 0.5 constant should be used as qreal(0.5) in the qRound.

So just to make sure, is your fix proposal following:
inline int qRound(qreal d) // here, qreal can be though of as float for ARM.

{ return d >= 0.0 ? int(d + qreal(0.5)) : int(d - int(d-1) + qreal(0.5)) + int(d-1); }

Comment 4 Liimatainen Jari Juhani (Nokia) error_managers 2011-01-13 10:17:17 EET

(In reply to comment #3)
> (In reply to comment #2)
> > You may also look at
> > https://projects.maemo.org/mediawiki/index.php?title=SystemAnalysis/NEON%2Bfloats
> > for other potential ways to handle these small Qt math related functions.
> >
> > As stated there the 0.5 constant should be used as qreal(0.5) in the qRound.
> So just to make sure, is your fix proposal following:
> inline int qRound(qreal d) // here, qreal can be though of as float for ARM.
>

{ return d >= 0.0 ? int(d + qreal(0.5)) : int(d - int(d-1) + qreal(0.5)) + > int(d-1); }

Yes, but it's good to verify with objdump still, is that enough, or should you
do explicit conversion also of 0.0.

I believe d-1 parts do not cause implicit double conversions, but that's also
good to veryfy with objdump so that the fix will be flawless.

For exact instructions how to verify it, please check the link provided in the
description!