PR57073 - Optimize __builtin_powif (-1.0, k) to k & 1 ? -1.0 : 1.0

[Tobias Burnus]

The attached patch optimizes
       __result_f = __builtin_powif (-1.0e+0, k);
to
   powi_cond_4 = k_1(D) & 1;
   powi_5 = powi_cond_4 ? -1.0e+0 : 1.0e+0;

I did an all-language (all,ada,go,objc++) bootstrap and regtest on 
x86-64-gnu-linux.
OK for the trunk?


Additionally, I would like to thank Thomas, who tried several approaches 
(e.g. fold_builtin_powi - which turned out to failed when called during 
optimization time, as regimplifing had problems with the COND_EXPR) and 
paved the way for this patch. And to Jakub, who gave helpful hints.

Note: Other replacements are also possible, e.g. "1.0 - (float) ((k & 1) 
<< 1)". However, the condition above turned out to be fasted as Jakub 
expected and Thomas has tested (see PR).


Tobias
-------------- next part --------------
A non-text attachment was scrubbed...
Name: powi.diff
Type: text/x-patch
Size: 2632 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20130530/661cad94/attachment.bin>