Re: [Qemu-devel] [PATCH 01/19] Add New softfloat Routines for VSX

[Alex Bennée]

address@hidden writes:

> On 25 October 2013 12:34, Alex Bennée <address@hidden> wrote:
>> Is it worth adding some sort of test into make check to defend these
>> softfloat functions against unintentional breakage? It would certainly
>> be worthwhile as soon as multiple arches use these functions as float
>> errors are often subtle and hard to track down.
>
> Ideally, but there's zero infrastructure for doing the kind
> of serious including-edge-cases testing at the moment, so I'm
> not really in favour of making it a gating condition for
> accepting patches.

I'm not proposing to halt inclusion for that I was just wondering aloud
how it could be defended. For the soft-float routines themselves they
could be tested within the existing tests/ stuff like
tests/check-qfloat.c without having to worry about hooking into target
arch specific test cases.

> If somebody wanted to set up such infrastructure, there are
> a couple of approaches that spring to mind:
>  (a) get risu (https://wiki.linaro.org/PeterMaydell/Risu) working
>   on more target architectures, add the "record-and-replay" feature
>   so it can be run without having target hardware, and then just
>   test softfloat by testing the actual target fp instructions

Interesting. Funnily we spent a lot of time at Transitive fixing up
translation failures that our random code generator threw up. It's also
equally interesting how far you can get with fairly broken translation
that no actual applications care about.

I'll have a look once I've fixed up build machinery around the existing
TCG tests.

>  (b) something involving wiring up IBM's IEEE test suite
>   vectors directly to our softfloat code:
>  
> https://www.research.ibm.com/cgi-bin/haifa/test_suite_download.pl?first=elenag&second=webmaster
>   (it's not clear to me what license the test vectors are
>   under)
>
> -- PMM

-- 
Alex Bennée