Re: GCC Status Report (2004-03-09)

[Ian Lance Taylor <ian@wasabisystems.com>]

Paul Koning <pkoning@equallogic.com> writes:

> >>>>> "Mark" == Mark Mitchell <mark@codesourcery.com> writes:
> 
>  Mark> The other alternative I can think of is to introduce a
>  Mark> scheduling barrier (e.g., a dummy vaolatile asm) after the
>  Mark> clearing operation. 
> 
> It would be nice if there were a clearly documented way to do that.
> 
> I've found two:
> 
>      asm("")
>      asm("":::"memory")
> 
> The former does what I want but isn't documented and (according to
> what I was told) should NOT be counted on to be a barrier.
> 
> The latter is documented and does a superset of what I needed at the
> time.  I say "superset" because it causes registers to be reloaded
> with data from variables (due to the memory clobber).  That's a
> performance hit if I didn't have any real memory clobbers in the
> system.
> 
> Perhaps we need asm("":::"barrier")?

Seconded.  For high performance device manipulation, it would be good
to have a way to say "do not schedule instructions across this point"
without also saying "clobber everything".

Ian