math-emu: Fix signalling of underflow and inexact while packing result.
authorKumar Gala <galak@kernel.crashing.org>
Wed, 22 Oct 2008 05:19:00 +0000 (22:19 -0700)
committerDavid S. Miller <davem@davemloft.net>
Thu, 23 Oct 2008 04:53:53 +0000 (21:53 -0700)
I'm trying to move the powerpc math-emu code to use the include/math-emu bits.

In doing so I've been using TestFloat to see how good or bad we are
doing.  For the most part the current math-emu code that PPC uses has
a number of issues that the code in include/math-emu seems to solve
(plus bugs we've had for ever that no one every realized).

Anyways, I've come across a case that we are flagging underflow and
inexact because we think we have a denormalized result from a double
precision divide:

000.FFFFFFFFFFFFF / 3FE.FFFFFFFFFFFFE
soft: 001.0000000000000 .....  syst: 001.0000000000000 ...ux

What it looks like is the results out of FP_DIV_D are:

D:
sign:   0
mantissa: 01000000 00000000
exp:  -1023 (0)

The problem seems like we aren't normalizing the result and bumping the exp.

Now that I'm digging into this a bit I'm thinking my issue has to do with
the fix DaveM put in place from back in Aug 2007 (commit
405849610fd96b4f34cd1875c4c033228fea6c0f):

[MATH-EMU]: Fix underflow exception reporting.

    2) we ended up rounding back up to normal (this is the case where
       we set the exponent to 1 and set the fraction to zero), this
       should set inexact too
...

    Another example, "0x0.0000000000001p-1022 / 16.0", should signal both
    inexact and underflow.  The cpu implementations and ieee1754
    literature is very clear about this.  This is case #2 above.

Here is the distilled glibc test case from Jakub Jelinek which prompted that
commit:

--------------------
#include <float.h>
#include <fenv.h>
#include <stdio.h>

volatile double d = DBL_MIN;
volatile double e = 0x0.0000000000001p-1022;
volatile double f = 16.0;
int
main (void)
{
  printf ("%x\n", fetestexcept (FE_UNDERFLOW));
  d /= f;
  printf ("%x\n", fetestexcept (FE_UNDERFLOW));
  e /= f;
  printf ("%x\n", fetestexcept (FE_UNDERFLOW));
  return 0;
}
--------------------

It looks like the case I have we are exact before rounding, but think it
looks like the rounding case since it appears as if "overflow is set".

000.FFFFFFFFFFFFF / 3FE.FFFFFFFFFFFFE = 001.0000000000000

I think the following adds the check for my case and still works for the
issue your commit was trying to resolve.

Signed-off-by: David S. Miller <davem@davemloft.net>
include/math-emu/op-common.h

index cc1ec396f8d61bcc6e82e851845184011d467eb6..bc50aa0d6d73c1a590b3cee3e56da2d2326d7491 100644 (file)
@@ -139,18 +139,27 @@ do {                                                              \
        if (X##_e <= _FP_WFRACBITS_##fs)                        \
          {                                                     \
            _FP_FRAC_SRS_##wc(X, X##_e, _FP_WFRACBITS_##fs);    \
-           _FP_ROUND(wc, X);                                   \
            if (_FP_FRAC_HIGH_##fs(X)                           \
                & (_FP_OVERFLOW_##fs >> 1))                     \
              {                                                 \
                X##_e = 1;                                      \
                _FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc);        \
-               FP_SET_EXCEPTION(FP_EX_INEXACT);                \
              }                                                 \
            else                                                \
              {                                                 \
-               X##_e = 0;                                      \
-               _FP_FRAC_SRL_##wc(X, _FP_WORKBITS);             \
+               _FP_ROUND(wc, X);                               \
+               if (_FP_FRAC_HIGH_##fs(X)                       \
+                  & (_FP_OVERFLOW_##fs >> 1))                  \
+                 {                                             \
+                   X##_e = 1;                                  \
+                   _FP_FRAC_SET_##wc(X, _FP_ZEROFRAC_##wc);    \
+                   FP_SET_EXCEPTION(FP_EX_INEXACT);            \
+                 }                                             \
+               else                                            \
+                 {                                             \
+                   X##_e = 0;                                  \
+                   _FP_FRAC_SRL_##wc(X, _FP_WORKBITS);         \
+                 }                                             \
              }                                                 \
            if ((FP_CUR_EXCEPTIONS & FP_EX_INEXACT) ||          \
                (FP_TRAPPING_EXCEPTIONS & FP_EX_UNDERFLOW))     \