Re: synching with minios

[Anil Madhavapeddy <anil@xxxxxxxxxx>]

On 11 Dec 2012, at 12:37, "Robert N. M. Watson" <robert.watson@xxxxxxxxxxxx> 
wrote:
> 
> On 11 Dec 2012, at 12:00, Balraj Singh wrote:
> 
>> Here and there, like Anil says is right.  Other than +., *., conversions and 
>> of course the maths functions, the only places that I found floating point 
>> (and sse) instructions when cursorily scanning the assembly was in the 
>> garbage collector and much of that use was for gc stats.  There is also a 
>> bit in sys.time.  I think that soft FP will not impact performance much if 
>> at all.  
> 
> However, I think what you might want is:
> 
> (1) Mirage parts don't require floating point -- ideally a compiler flag 
> combination for those portions to say "If it implies FP code generation, fail"
> (2) Runtime and core OCaml libraries that can be linked against soft FP if 
> desired, but can also use native FP if desired
> (3) The ability for applications linked against Mirage to use FP natively for 
> performance reason ... if desired.
> 
> That is: reduce or eliminate dependence on FP at the lowest levels of the 
> system, but permit its use higher up if applications want it. If you're doing 
> some FP-exploiting thing at the application level, definitely don't want to 
> forbid its use of hardware FP!

Agreed.  A quick glance at GCC implies that it doesn't like doing soft-float on 
x86 very much either, so starting there at the runtime level and working our 
way up makes sense.

I'm deferring this one until we start upstreaming the kFreeBSD backend as it 
looks like we'll need a robust cross-compilation toolchain for both GCC and 
OCaml.

-anil