On Fri, Nov 13, 2015 at 12:27 PM, Richard Sandiford
<richard.sandif...@arm.com> wrote:
> Richard Biener <richard.guent...@gmail.com> writes:
>> On Tue, Nov 10, 2015 at 1:09 AM, Bernd Schmidt <bschm...@redhat.com> wrote:
>>> On 11/07/2015 01:46 PM, Richard Sandiford wrote:
>>>>
>>>> @@ -3814,8 +3817,8 @@ extract_range_basic (value_range *vr, gimple *stmt)
>>>> break;
>>>> /* Both __builtin_ffs* and __builtin_popcount return
>>>> [0, prec]. */
>>>> - CASE_INT_FN (BUILT_IN_FFS):
>>>> - CASE_INT_FN (BUILT_IN_POPCOUNT):
>>>> + CASE_CFN_FFS:
>>>> + CASE_CFN_POPCOUNT:
>>>> arg = gimple_call_arg (stmt, 0);
>>>> prec = TYPE_PRECISION (TREE_TYPE (arg));
>>>> mini = 0;
>>>
>>>
>>> So let me see if I understood this. From what we discussed the purpose of
>>> these new internal functions is that they can have vector types. If so,
>>> isn't this code (here and elsewhere) which expects integers potentially
>>> going to be confused?
>>
>> We indeed need to add additional checks to most users of CASE_CFN_* to cover
>> the bigger freedom that exists with respect to types.
>
> The code above is OK because it's only handling integral types.
> A vector popcount or vector ffs must return a vector result.
>
>> Richard, please audit all the cases you change for that.
>
> I had another look and the only problematical uses I could see are the
> match.pd ones. E.g.:
>
> /* Optimize logN(func()) for various exponential functions. We
> want to determine the value "x" and the power "exponent" in
> order to transform logN(x**exponent) into exponent*logN(x). */
> (for logs (LOG LOG LOG LOG2 LOG2 LOG2 LOG10 LOG10)
> exps (EXP2 EXP10 POW10 EXP EXP10 POW10 EXP EXP2)
> (simplify
> (logs (exps @0))
> (with {
> tree x;
> switch (exps)
> {
> CASE_CFN_EXP:
> /* Prepare to do logN(exp(exponent)) -> exponent*logN(e). */
> x = build_real_truncate (type, dconst_e ());
> break;
> CASE_CFN_EXP2:
> /* Prepare to do logN(exp2(exponent)) -> exponent*logN(2). */
> x = build_real (type, dconst2);
> break;
> CASE_CFN_EXP10:
> CASE_CFN_POW10:
> /* Prepare to do logN(exp10(exponent)) -> exponent*logN(10). */
> {
> REAL_VALUE_TYPE dconst10;
> real_from_integer (&dconst10, VOIDmode, 10, SIGNED);
> x = build_real (type, dconst10);
> }
> break;
> default:
> gcc_unreachable ();
> }
> }
> (mult (logs { x; }) @0))))
>
> Here we could either add a SCALAR_FLOAT_TYPE_P check or extend
> build_real to handle vector types. Which do you think would be best?
I don't like extending build_real. For now add a SCALAR_FLOAT_TYPE_P.
Or maybe extend build_real ...
Well. Go with SCALAR_FLOAT_TYPE_P.
Thanks,
Richard.
>
> Thanks,
> Richard
>
