------- Additional Comments From jbeitaharon at intrusic dot com 2005-05-19
20:43 -------
Subject: Re: Optimization level -O2 breaks float (double) conversion
Dear Pinskia and Co.,
As a former compiler developer myself I was just trying to help, and I'm
sorry to tell you that your answer was not to the point.
In order to transmit this double-precision-float we have to reformat it
into/from "Network Byte Ordering". The compiler generated no warnings,
and the generated code works fine when un-optimized or when optimized
with -O1.
The problem reproduction program reformats X into Y and then reformats Y
into X with the inverse operation. It then verifies that the original X
has the same value as the final X.
There is absolutely no reason why the -O2 optimization should break the
code, and hence this bug is valid as a regression in -O2 relative to -O1.
Regards,
Jonathan
pinskia at gcc dot gnu dot org wrote:
>------- Additional Comments From pinskia at gcc dot gnu dot org 2005-05-19
>19:15 -------
>You are violating C aliasing rules:
>((long *) &s)[0] = __extension__ ({ register u_long __X = (nt.li[1]); __asm
>("xchgb %h1, %b1\n\trorl $16, %1\n\txchgb %h1, %b1" : "=q" (__X) : "0" (__X));
>__X; });
>
>s is a double but you are accessing it as a long.
>
>
>
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21666
