[dragonegg PATCH] Implement __builtin_signbit

[Peter Collingbourne]

On Thu, Jun 27, 2013 at 08:51:21PM +0200, Duncan Sands wrote:
> Hi Peter,
>
> >>> +      return Builder.CreateZExt(Result,
>>>> +                                ConvertType(gimple_call_return_type(stmt)));
>>>
>>> I think the extension type should be based on the signedness of the return type,
>>> rather than assuming it is an unsigned type.
>>
>> Result is an i1, so zero extension should cause us to return 0 or 1
>> (consistent with GCC) regardless of the signedness of the return type.
>
> it does look like GCC only cares about 0 or no-zero.  In which case I'm OK with
> this.  That said, it could be more efficient.  For example, if the input type is
> float, and the output i32, then you can just and with a mask that only keeps the
> sign bit, reducing from bitcast+icmp+zext to bitcast+and.  What do you think of
> doing a bitcast, an "and" to only keep the sign bit, then
>   - if the result is wider than the float, a zext
>   - if the result is the same size as the float, nothing
>   - if the result is smaller than the float, right shift the bit into the 
> top bit of the result

'd be wary of making such a change, not only to stay consistent with
GCC and Clang, but because IR instruction count isn't necessarily
correlated with machine instruction count.  In particular, x86 has
the MOVMSKPD and MOVMSKPS instructions which can be selected for the
code this function generates while it would take another instruction
for the scheme you describe.  I also don't think this intrinsic will
come up often enough for us to care too much about performance so
I'm tempted to leave this as is.

Thanks,
-- 
Peter