This is the mail archive of the libc-help@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Switch implementation based on circumstances.


On Thu, Dec 06, 2012 at 12:44:04PM -0600, Ryan S. Arnold wrote:
> On Thu, Dec 6, 2012 at 12:40 PM, OndÅej BÃlka <neleai@seznam.cz> wrote:
> > Hi,
> >
> > I noticed that at several places we can use fast implementation most of
> > time and switch to more generic one only after some function is called.
> 
> Could you use the IFUNC mechanism for this?
> 
> https://www.ibm.com/developerworks/wikis/display/LinuxP/Optimized+Libraries#OptimizedLibraries-TargetSpecific
>
Not exactly. I could add there a hook to change implementations when
relevant functions will be called.
 
> > This mostly applies to fact that in single thread programs locking is
> > not needed. I want to do something like compile file twice first time
> > with locks on and second off. Then I would and hook in pthread_create+shmat
> > that switches implementation.
> >
> > malloc, printf ... would benefit most from this one.
> 
> 
> The problem with having a single threaded application traverse down a
> non-locking path in code like malloc is that there are always the
> cases of signal handler abuse.  Someone calls a function in a signal
> handler that calls malloc.  But malloc was interrupted in mid-process.
> So instead of getting deadlock like what happens now, you get a huge
> undebuggable mess.
> 

What about swithching implementation also for sigaction?

> 
> I'm in the camp that I don't think printf should ever been seen as a
> performance path that needs improvement.  The runtime type checking
> requirements introduce way too many burdens.
> 
> 
> > Second is locale depended optimalizations where a hook for setlocale is
> > needed.
> >
> > I do not know what is best way to collect hooks. One possibility is single
> > file where you write all affected functions.
> 
> More opportunity for IFUNC.
> 
> I think when implementing some new features and writing new code that
> using STM might be better than trying to have two implementations.
> 
> Ryan

-- 

You need to upgrade your VESA local bus to a MasterCard local bus.


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