[Patch, fortran] PR32239 Use builtin_powi
Janne Blomqvist
blomqvist.janne@gmail.com
Sun Jul 1 17:19:00 GMT 2007
Janne Blomqvist wrote:
> On 6/27/07, Steve Kargl <sgk@troutmask.apl.washington.edu> wrote:
>> On Wed, Jun 27, 2007 at 11:23:19AM +0200, Tobias Burnus wrote:
>> > Tobias Burnus wrote:
>> > > Ok for the trunk. I assume you have regtested it; in any case it
>> builds
>> > > and some limited testing shows no problems.
>> > >
>> > I bootstrapped and regtested it (x86-64/Linux). For -m32 everything is
>> > ok, but for -m64 it fails for:
>> >
>> > gfortran.fortran-torture/execute/intrinsic_fraction_exponent.f90's
>> > test_4 (third call).
>> >
>> > Instead of:
>> > test_4; x = 1.4012985E-45 y= 0.5000000 z= -148
>> > test_4; y= 1.4012985E-45 |x-y|= 0.000000 |x*1e-6|= 0.000000
>> > the result is:
>> > test_4; x = 1.4012985E-45 y= 0.5000000 z= -148
>> > test_4; y= 0.000000 |x-y|= 1.4012985E-45 |x*1e-6|= 0.000000
>> > and the latter fails because |x-y| > |x*1e-6|.
>> >
>> > The dump looks ok:
>> > - y = (y * 2.0e+0) * _gfortran_pow_r4_i4 (2.0e+0, z + -1);
>> > + y = (y * 2.0e+0) * __builtin_powif (2.0e+0, z + -1);
>> > }
>> > else
>> > {
>> > - y = (y / 2.0e+0) * _gfortran_pow_r4_i4 (2.0e+0, z + 1);
>> > + y = (y / 2.0e+0) * __builtin_powif (2.0e+0, z + 1);
>> >
>> > Thus the builtin_powi gives a different result; I don't see ad hoc
>> > whether both results are ok or whether one implementation has a bug.
>> >
>>
>> Looks like an issue with subnormal numbers. If 1.4012985E-45 isn't
>> the smallest subnorml number, it's awefully close. y/2. has underflowed
>> to zero.
>
> The algorithms implemented by _gfortran_pow vs. builtin_powi
> (builtin_powi in the PR) basically do the same thing, I don't think we
> can say one is better than the other.
>
> Perhaps the testcase should be changed to test something like
>
> abs(x-y) < ulp * spacing(max(abs(x),abs(y)))
>
> instead of the somewhat ad hoc test used now?
After a brief discussion on irc, committed to trunk along with the
testsuite fix.
--
Janne Blomqvist
More information about the Fortran
mailing list