Re: [Qemu-devel] ppc: TCG and FP exceptions, is it right ?

[Tristan Gingold]

> On 25 Jul 2016, at 11:44, Benjamin Herrenschmidt <address@hidden> wrote:
> 
> Hi folks !
> 
> I could use a bit of help determining if something is quite right in
> TCG emulation of FP ops.
> 
> Fabien, Tristan, you submitted a patch back in 2013:
> 
> db72c9f256ae70b30c5d5985234f085df4226c55 "powerpc: correctly handle fpu
> exceptions" which effectively makes any of the exceptions generated
> for underflow, overflow and inexact fire right away
> (helper_raise_exception_err will exit the cpu loop with a setjmp).
> 
> This makes them effectively do the same thing
> as float_zero_divide_excp().
> 
> However I *think* it might not be what we want, according to the
> comment in helper_float_check_status() that says
> 
>  /* Differred floating-point exception after target FPR update */
> 
> And according to the architecture definition for those exceptions,
> where we indeed want the target FPR updated before we take the
> interrupts.
> 
> The code as writte will take the exception before the FPR is updated,
> I *think*, or am I missing something here  ?

This looks like an oversight. In our tests, we don't read the fpr.

> 
> I think the intent was to return to the translated code so the FPRons
> update happen, though we ideally would need to also set some state
> so the translated code itself can then check for an exception and
> fire it.
> 
> However as you noticed, that doesn't work well either.
> 
> What do you think is the most appropriate implementation here?
> 
> I'm thinking it's almost worth bringing FE0/FE1 into the hflags
> so that we know at translation time whether to be precise, imprecise,
> or ignore FP exceptions.
> 
> Then we could do something along the lines of:
>  
>   - In the helpers, when checking status, set an env flag if an
> exception should occur.
> 
>   - In all the translate call sites, if FE0/1 is non-0 (at translate
> time), generate call to check that flag and shoot the exception
> 
>   - Optionally, we could even implement some smarts to defer this to
> the end of a TB in imprecise mode.
> 
> An additional note is that if FE0/FE1 are 0, we still in some case
> leave an exception behind in cs->exception_index. Now, I *think*
> that's ok, it will just be silently dropped at some point, but I am not
> 100% certain as that's a part of TCG I'm an not super familiar with
> yet.
> 
> What do you guys reckon ? I am missing something here ?

I like the second helper suggestion from address@hidden  That's simple.

Tristan.