This is the mail archive of the mailing list for the GNU Fortran project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [patch, libgfortran] PR48906 Wrong rounding results with -m32

On 06/14/2011 12:18 AM, Thomas Henlich wrote:
On Tue, Jun 14, 2011 at 06:51, jerry DeLisle<> wrote:
It should be easy to implement:

After the switch between F and E editing, we just need to shift the
decimal point and decrement the exponent. No new rounding is required,
because we keep the number of significant digits.

OK, after a little bit of experimentation, I have arrived at the updated patch attached.

This has been regression tested and passes all test cases I am aware of.  I
also have included a new test case gcc/testsuite/gfortran.dg/fmt_g.f90.

OK for trunk?

I have reviewed your patch, and I noticed that you placed the digit-shifting code quite at the top of output_float(), where the final value of e is not even known. Due to rounding, e can be modified after this point, so your code will generate invalid output in some cases, for example:

print "(-2PG0)", nearest(0.1d0, -1.0d0) ! 1.0000000000000000E+001
expected .0099999999999999992E+001

Please put the code where at belongs, after the switch between F and E
editing (based on the final value of e).

The same applies to the scale factor in general, e.g.

print "(-2pg12.3)", 0.096    ! 1.00E+01 expected 0.001E+02
print "(-1pg12.3)", 0.0996   ! 1.00E+00 expected 0.010E+01
print "(-2pg12.3)", 0.09996  ! 1.00E+01 expected 0.100
print "(-1pg12.3)", 0.09996  ! 1.00E+00 expected 0.100
print "(1pg12.3)",  0.099996 ! 1.000E-01 expected 0.100
print "(2pg12.3)",  0.099996 ! 10.00E-02 expected 0.100
print "(-2pg12.3)",  999.6  ! 0.100E+04 expected 0.001E+06
print "(-1pg12.3)",  999.6  ! 0.100E+04 expected 0.010E+05
print "(1pg12.3)",  999.6  ! 0.100E+04 expected 9.996E+02
print "(2pg12.3)",  999.6  ! 0.100E+04 expected 99.96E+01

Please revise your code to fix this. A working approach I have outlined in
and an (alpha) implementation is here:

Yes, I tried the alpha approach and got quite a number of regressions in the test suite. This lead me to think it would be better to start at the point of fewest regressions and work from there.

Regardless, continued patience is called for on this and I will keep at it as time permits.



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]