On 05/16/2017 09:54 AM, David Hildenbrand wrote:
> On 15.05.2017 11:19, Christian Borntraeger wrote:
>> Nice to see somebody doing this for s390x.
>>
>>
>> On 05/12/2017 12:58 PM, David Hildenbrand wrote:
>> [...]
>>
>>> diff --git a/s390x/Makefile b/s390x/Makefile
>>> new file mode 100644
>>> index 0000000..f9468bb
>>> --- /dev/null
>>> +++ b/s390x/Makefile
>>> @@ -0,0 +1,32 @@
>>> +all: test_cases
>>> +
>>> +test_cases: $(tests)
>>> +
>>> +CFLAGS += -std=gnu99
>>> +CFLAGS += -ffreestanding
>>> +CFLAGS += -Wextra
>>> +CFLAGS += -I lib
>>> +CFLAGS += -O2
>>> +CFLAGS += -msoft-float
>>
>>
>> If you enable the AFP register control bit in your startup code, you can probably
>> avoid the soft-float thing.
>>
>
> Looks like we can drop that completely.
>
> a) no floating point operations are used either way.
>
> b) adding a very basic division results in
>
> static void test_fp(void)
> {
> double a, b;
>
> b = 6.5;
> a = 2.0;
>
> b = b / a;
>
> report("test fp", b == 3.25);
> }
>
> (compiled with -O0)
>
> s390x/selftest.o: In function `test_fp':
> /home/dhildenb/git/kvm-unit-tests/s390x/selftest.c:21: undefined
> reference to `__divdf3'
> /home/dhildenb/git/kvm-unit-tests/s390x/selftest.c:23: undefined
> reference to `__eqdf2'
>
> So we would need a soft float lib (as we're using -nostdlib)
>
> c) dropping -msoft-float
>
> A DDB (DIVIDE long BFP) is generated. That seems to work under TCG
> without any exception. I assume the check for AFP-register control is
> missing. I assume, that test would not pass under KVM.
>
>
> For now I will simply drop -msoft-float.
This will allow gcc to spill/fill into floating point registers, which are
not all available without afp-register control. It will probably work because
the gr->fpr move will not be used for -march=z900, but I suggest to not rely
on that.
