Re: [PATCH 2/8] tracing: create automated trace defines

[Mathieu Desnoyers <compudj@krystal.dyndns.org>]

* Steven Rostedt (rostedt@goodmis.org) wrote:
> 
> [ removed Pekka@firstfloor.org due to mail errors ]
> 
> On Tue, 21 Apr 2009, Frank Ch. Eigler wrote:
> 
> > Hi -
> > 
> > On Tue, Apr 21, 2009 at 05:17:17PM -0400, Steven Rostedt wrote:
> > 
> > > [...]  Perhaps we should fork off gcc and ship Linux with its own
> > > compiler. This way we can optimize it for the kernel and not worry
> > > about any userland optimizations.
> > 
> > In this regard, kernel land does not seem that unlike user land.
> > 
> > > 	if (unlikely(err)) {
> > > 		__section__(".error_sect") {
> > > 			/* put error code here */
> > > 		}
> > > 	}
> > > 
> > > And have gcc in the error section (if it is big enough perhaps) do:
> > > 	jmp .L123
> > > .L124   [...]
> > > [...]
> > > 	jmp .L124
> > 
> > > We could do the same for trace points. That is, any part of code that 
> > > really would happen once in a while (error handling for one) we can move 
> > > off to its own section and keep hot paths hot.
> > 
> > This is called -freorder-blocks or -freorder-blocks-and-partition
> > (depending on how far you would like gcc to move unlikely blocks).
> 
> That does not let us pick and choose what and where to put the code.
> 
> But still, a fork of gcc would let us optimize it for the kernel, and not 
> for generic programs.
> 
> /me has been sitting too close to the furnace and must have been taking 
> up some of those fumes, to be considering a fork of gcc a good idea ;-)
> 

I guess we should have been sitting near the same furnace then. I'm
unsure how different from the current gcc this can go, but it could be a
very interesting exercise. Just removing unneeded front ends could
probably help adding features much faster than if we have to support
Fortran, Java, etc.

Mathieu

> -- Steve
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68