Re: [RFC PATCH 1/2] target/ppc: Add invalid imz, isi and snan to do_float_check_status()

[Richard Henderson]

On 5/12/22 12:42, Víctor Colombo wrote:
> Make do_float_check_status also handle some invalid exceptions:
> imz, isi and snan.
>
> The idea is to try to concentrate the handling of all fp exceptions in
> do_float_check_status. For this RFC, move only the 3 exceptions stated
> above.
>
> I chose to create new '*_defer' helpers to propose a gradual transition
> for the instructions helpers. The 'defer' comes from the idea that the
> exception raising is defered to be called somewhere else, as oposed of
> what is being done in the original, 'non-defered' functions, which
> immediatly call raise_exception_err_ra().

This is not correct.

You need to go through section 4.4 "Floating-point excepions" and note what happens when 
each exception is enabled (e.g. VE=1):

(1) Do not modify the FPR,
(2) Do or do not modify FR, FI, FPRF.

The mechanism by which we avoid setting the FPR is to to raise an exception before 
returning from the helper performing the operation.

I believe it's mostly invalid and divbyzero that avoid setting results, while overflow, 
underflow, inexact do.


r~