autoconf-patches
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: changing "configure" to default to "gcc -g -O2 -fwrapv ..."


From: Andreas Büning
Subject: Re: changing "configure" to default to "gcc -g -O2 -fwrapv ..."
Date: Sat, 30 Dec 2006 12:27:09 +0100
User-agent: Mozilla/5.0 (OS/2; U; Warp 4.5; de-DE; rv:1.3) Gecko/20030616

Richard B. Kreckel wrote:

Bottom line: without such a warning, -fwrapv should be the default and
should not be turned off by any -O option.

I agree with this. At least until -O2 no optimizations should be done
by default if they can modify the result of an operation even if this
might be a naive compiler user's assumption. As a "naive" user I know
that without optimization compilers may output partially "strange" code.
I remember a case (long time ago) when gcc without optimization produced
something like mov ax, bx; mov bx, ax; which is complete nonsense because
the 2nd operation is redundant. But that's ok because it's a side effect
when the compiler doesn't search for redundant code. But for me this means
I have to turn on some optimization to get reasonable code. And I expect that
there is an optimization switch for "remove all redundant code and do all
possible optimizations without changing the program logics". -O2 seemed to
be quite common for this among different compilers. With -O3 I expect 
"modifying"
optimizations to be done which means I have to be _really_ careful whether
I can enable that optimization level. This is also ok. But if I can't rely
on -O2 to be "safe" then I might be forced to disable optimization which also
might mean that I should change the compiler.

Please, don't get me wrong. I'm not sure how important this -fwrapv feature
is to real world code but it sounds dangerous to me to do this optimization
with -O2. And from my naive point of view I would expect that it would be
enabled with -O3 or above, not with -O2. I personally do not care much
about benchmark optimization because neither do I develop a benchmark nor
do I use a benchmark for my daily work. And I think it's the same for 99%
of all gcc users. If you _really_ need one single flag to optimize your
benchmarks then introduce an -Obenchmark flag or so but, please, don't pollute
-O2 with stuff that might produce highly optimized but not 100% reliable code.


Andreas





reply via email to

[Prev in Thread] Current Thread [Next in Thread]