On 06/16/2011 06:25 PM, Richard Henderson wrote:
> On 06/16/2011 05:44 AM, Bernd Schmidt wrote:
>> +@deftypefn {Built-in Function} int __builtin_clrsb (unsigned int x)
>> +Returns the number of leading redundant sign bits in @var{x}, starting
>> +at the most significant bit position.
>> +@end deftypefn
> 
> Do we want a signed argument, since we're talking about signs?

Err, yes. It's signed everywhere else (builtins.def etc.).

> It would seem that unlike clz, this function is not undefined for zero.
> What about INT_MIN?  Do all cpus handle those edge cases the same way?

-1 and zero should both produce the same value, 31 (for a 32 bit
integer). I don't see why INT_MIN should be special - the return value
is zero. This is true for C6X and Blackfin; ARM documentation suggests
it's also true for their VCLS instruction. I've not found proper
picochip documentation but some other documents that suggest it's also
implemented this way.

> Do you get smaller code in general from
> 
>   if (x < 0)
>     x = ~x;
>   if (x == 0)
>     return W_TYPE_SIZE - 1;
>   count_leading_zeros(ret, x);
>   return ret - 1;

Probably.

>> -(define_insn "signbitssi2"
>> +(define_insn "clrsbsi2"
>>    [(set (match_operand:HI 0 "register_operand" "=d")
>>  	(if_then_else:HI
>>  	 (lt (match_operand:SI 1 "register_operand" "d") (const_int 0))
> 
> No use of the new rtx code?

D'oh. Blackfin has a (clrsb:HI (operand:SI)) instruction, so adding this
showed a problem with some of the existing simplify_const_unop cases:
for ffs/clz/ctz/clrsb/parity/popcount, we should look at the mode of the
operand, rather than the mode of the operation. This limits what we can
do in that function, since op_mode is sometimes VOIDmode - we really
should add builtin folders for these at some point.

New patch below. Retested on i686 and bfin.


Bernd