On 12/06/2019 20:45, Richard Henderson wrote:
> On 6/2/19 4:08 AM, Mark Cave-Ayland wrote:
>> void helper_xvxsigsp(CPUPPCState *env, uint32_t opcode)
>> {
>> - ppc_vsr_t xt, xb;
>> + ppc_vsr_t *xt = &env->vsr[xT(opcode)];
>> + ppc_vsr_t *xb = &env->vsr[xB(opcode)];
>> + ppc_vsr_t t = *xt;
>> uint32_t exp, i, fraction;
>>
>> - getVSR(xB(opcode), &xb, env);
>> - memset(&xt, 0, sizeof(xt));
>
> Change in behaviour -- zero init to copy init.
>
> Note for future cleanup: most of these initializations do not need to happen,
> because we overwrite all elements of T without consuming the previous value.
>
>
>> @@ -3410,23 +3382,22 @@ void helper_xsrqpi(CPUPPCState *env, uint32_t opcode)
>> env->fp_status.float_exception_flags &= ~float_flag_inexact;
>> }
>>
>> - helper_compute_fprf_float128(env, xt.f128);
>> + helper_compute_fprf_float128(env, t.f128);
>> + *xt = t;
>> do_float_check_status(env, GETPC());
>> - putVSR(rD(opcode) + 32, &xt, env);
>
> Change in behaviour -- writeback happens before do_float_check_status instead
> of after. This may well be a bug fix, but if so should happen separately.
I've now fixed both of these. FWIW this is probably going to be my last bit of
work
on VSX for a while since as I don't have 64-bit reference hardware, and writing
and
testing these patchsets takes a long time. At least I feel that things are now
in a
place where people can start to take more advantage of the vector ops should
they wish.
ATB,
Mark.
