Re: loader breaks with -O2 optimizations

Stephan F. Yaraghchi wrote:
> On 10/31/07, Kris Kennaway <kris_at_freebsd.org> wrote:
>> Stephan F. Yaraghchi wrote:
>>> On 10/31/07, Kris Kennaway <kris_at_freebsd.org> wrote:
>>>> Stephan F. Yaraghchi wrote:
>>>>> After making world on a freshly installed 7.0-BETA1
>>>>> the system does not boot anymore due to a broken loader:
>>>>>
>>>>>
>>>>> FreeBSD/i386 bootstrap loader, Revision 1.1
>>>>> (root_at_bigblue, Tue Oct 30 11:26:32 CET 2007)
>>>>> Can't work out which disk we are booting from.
>>>>> Guessed BIOS device 0xffffffff not found by probes defaulting to disk0:
>>>>>
>>>>> panic: free: guard1 fail_at_ 0x6ded4 from
>>>>> /usr/src/sys/boot/i386/loader/../../common/module.c:959
>>>>>
>>>>> --> Press a key on the console to reboot <--
>>>>>
>>>>>
>>>>> I found out that the following line in my make.conf causes
>>>>> the problem:
>>>>>
>>>>> CFLAGS= -O2 -funroll-loops -pipe
>>>>>
>>>>>
>>>>> After changing down to -O1 and making /usr/src/sys/boot again
>>>>> the systems behaves properly at boot.
>>>>>
>>>>> Is this behavior intended?
>>>> Maybe, what happens if you use just -O2 -pipe? -funroll-loops is not an
>>>> appropriate thing to be using globally anyway, unless your intention is
>>>> to randomly make some code slower.
>>>>
>>>> Kris
>>>>
>>> Hi Kris,
>>>
>>> I tried all possible combinations of these switches -- only -O2 led to
>>> the described
>>> behaviour.
>> Presumably you mean -O2 -funroll-loops, not -O2.  Or are you saying the
>> latter also breaks the loader?
>>
>>> Anyway, it's very interesting to hear that adding these optimizations
>>> to make.conf
>>> is not recommended, even that -funroll-loops is possibly slowing down
>>> certain code.
>>>
>>> I'm sure many people use it since it's a common tuning tip found on the web.
>>> I read about it in Dru Lavigne's "BSD Hacks" (O'Reilly)...
>> Yes, unfortunately it's bogus advice.  What does she say this option is
>> good for?
>>
>> Kris
>>
> 
> Hi Kris,
> 
> you are right: It's the combination of both that causes loader to break.
> 
> Oliver suggested the defaults (-O2 -fno-strict-aliasing -pipe) and it works.
> 
> By the way: "BSD Hacks" is a collection of hacks compiled by Dru.
> Hack #69 deals with tuning and is authored by Avleen Vig.
> It says: "The -funroll-loops saves one CPU register that would otherwise
> be tied up in tracking the iteration of the loop, but at the expense of making
> a slightly larger binary."

OK, so it's factually correct but just not good advice :)

Kris