Re: [avr-gcc-list] Avr-gcc Produces Incorrect Code with -Os

[John Regehr]

> > # define sei()  __asm__ __volatile__ ("sei" ::: "memory")
> > # define cli()  __asm__ __volatile__ ("cli" ::: "memory")
>
> This has been discussed before (on the avr-libc list, December 2005, just 
> search for '"cli" and "sei" should clobber memory'), and this exact solution 
> was actually proposed by me.

Adding these clobbers is an extremely good idea because they make cli/sei 
act like proper locks, in two senses.  First, they prevent reordering of 
computations as has been discussed here.  Second, they force data cached 
in registers to be written to RAM when the interrupt lock is released. 
Both of these are important.

Here's a subtle but important point: If you consistently protect shared 
data with clobbering locks, you don't need to make data volatile when it 
is shared between interrupts and main().  Of course you still need 
volatile for accessing hardware registers.

Although making the odd byte-sized flag volatile in order to support 
synchronization is not a bad idea (and of course we all do it), volatile 
as a basis for synchronization scales up to large programs with 
interesting shared data structures extremely poorly.  Better we should all 
use proper locks.

> The downside is that a memory clobber like that drops all optimizations of 
> global variables that the compiler might be temporarily storing in registers. 
> So, it actually hurts performance for code unrelated to the critical section.

We studied the impact of adding clobbers to critical sections in TinyOS 
and found that both code size and CPU usage actually decreased slightly 
due to the clobbers.  No idea why.  Details here:

  http://www.cs.utah.edu/~regehr/papers/plos06b.pdf

TinyOS currently clobbers memory on every cli/sei.

John Regehr