This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: /internet




> If I understand correctly, Jeff's proposal was to respect ordering when
> the user provides parentheses, but to leave the compiler free to do
> reordering when the user does not provide parentheses.

That would lead immediately to getting different answers on different
machines, or even different answers after slight code changes on the
same machine.  The order of evaluation would be changed to suit each
particular machine's scheduling algorithm.  It should not be permitted
as the normal behavior.

Since reordering can produce infinitely large numerical changes,
I have argued against allowing associative law optimization even
in fast-math mode.  It ought to be well isolated with its own
command switch, -fvectorize or some such.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]