[patch] Shark reroute LLVM atomic intrinsics to Zero

[Gary Benson]

Matthew Flaschen wrote:
> Gary Benson wrote:
> > Having said that, is there any reason to use atomic intrinsics on a
> > single processor system?  And, do multiprocessor ARM systems exist?
> 
> I think the answers are yes, and definitely yes.  Even with a single
> core, a thread can still be preempted.  And there are definitely
> multi-processor ARM
> (http://www.arm.com/products/CPUs/ARMCortex-A9_MPCore.html).

Ok, so there's no way around this.

I'm thinking is that ARM is a special case here: you could implement
SharkBuilder::CreateCmpxchg* on ARM by inlining arm_compare_and_swap
from atomic_linux_zero.inline.hpp.  That wouldn't require any LLVM
modifications, and it wouldn't be any slower than a "real"
implementation of llvm.atomic.cmp.swap.i32 because either way the
atomic part is handled by the call to __kernel_cmpxchg so the
generated code would be the same.

Obviously it would be preferable for this code to be in LLVM rather
than in Shark, but I know that implementing it in LLVM won't be quick
or easy so I can live with a bit of clutter in Shark for this.

What I'm not happy with is the idea of a generic solution that would
perform a function call on platforms that don't require one, because
that _would_ be slower than an implementation of llvm.atomic.cmp.swap.
For platforms that have instructions to do this I'm going to insist
people bite the bullet and implement the intrinsic.

Cheers,
Gary

-- 
http://gbenson.net/