Re: RFC: Preserving order of functions and top-level asms via cgraph

[Mark Mitchell <mark at codesourcery dot com>]

Daniel Jacobowitz wrote:
> On Sun, Apr 10, 2005 at 05:56:55PM -0700, Mark Mitchell wrote:
>
> > I don't really know how to prove the theory that we can lose the 
> > functionality.  I can see how we can disprove the theory: someone can 
> > present an example (as you did), for which we develop the consensus 
> > "yes, that's a reasonable thing for a user to do, so we must support 
> > this."  I suppose we could hardwire -funit-at-a-time to on for a while 
> > (as it happens to break the ordering), and see how many people yell at 
> > us, and then, if there was no persuasive yelling, consider that we've 
> > proved the theory.
>
>
> It's on by default; it has been for a long time.  People explicitly
> turn it off, i.e. there are users of the -fno-unit-at-a-time option.
> I think that qualifies as proof by yelling.

What I'd like to know is why these users are using the option, so that 
we can see if we think those uses are reasonable.

If we simply take the position that we can't break existing code, then 
we certainly can't do what I'm suggesting.  In general, we certainly 
don't want to break existing code, but maybe if we look at the examples, 
we'll decide we don't mind.  On the other hand, maybe we'll decide 
they're valid.  Or, maybe we'll decide that there are so many examples 
that breaking the code is a bad idea, even if each individual example is 
a disgusting hack.

There might also be better ways to do the things that people are 
currently doing, so that we could break existing code, but provide a 
useful workaround.  For example, we could add an asm_before function 
attribute that contained literal assembly code to be emitted directly 
before the function.  That might serve the MIPS kernel needs, in a 
cleaner way than the current implicit ordering constraint.  A further 
advantage would be that if, say, the compiler could prove the function 
unused (which will not be the case in the situation you describe), then 
the assembly code would never be emitted.  Right now, we might get in 
trouble if the assembly fragment were emitted, but the function elided.

--
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304