Re: [Qemu-devel] [PATCH] rsqrte_f32: No need to copy sign bit.

[Christophe Lyon <christophe.lyon@st.com>]

On 09.10.2011 00:57, Peter Maydell wrote:
> On 3 October 2011 15:28, Christophe Lyon<christophe.lyon@st.com>  wrote:
>> Indeed, the result is known to be always positive.
>> -    val = ((val64>>  63)&  0x80000000)
>> -        | ((result_exp&  0xff)<<  23)
>> +    val = ((result_exp&  0xff)<<  23)
>>          | ((val64>>  29)&  0x7fffff);
>>      return make_float32(val);
> So we weren't generating incorrect results, we were just doing
> slightly more work than we really needed, right? I'm curious
> what prompted this patch :-)
>
Exactly. And no way to expose a bug :-(

I was reading 2 revisions of the ARM ARM, and noticed erratas in the descriptions of FPRecipEstimate and FPRSqrtEstimate.

Sign propagation has been removed in the former, and not in the later, so I re-read both functions carefully as well as qemu's implementation and came to this conclusion :-)

I have contacted ARM support and suggested them to fix both function accordingly :-)

Christophe.