Bug in valarith.c:value_equal()?

Bug in valarith.c:value_equal()?

[Mark Kettenis]

I've found the cause of the testsuite problems I reported yesterday.
The additional testsuite failures are intermittent.  If you look
careful at the gdb.mi/mi-var-cmd.exp test you'll see that the test is
checking whether some uninitialized local variables have been changed.
The testsuite failures indicate that sometimes, the floating-point
variables change unexpectedly.  Some further investigation showed that
these unexpected changes happened when the (unitialized) variables
were NaNs.  All of a sudden things make sense.  The variables don't
really change.  GDB tries to determine whether a variable changes by
comparing its current value to a previous value.  This is done by
calling valarith.c:value_equal().  For floating-point variables, this
function does the following check:

    return value_as_double (arg1) == value_as_double (arg2);

Now in C this will return 0, if ARG1 and ARG2 are NaN, even if they
are bit for bit equal.

Actually I think the implementation of valarithm.c:value_equal() is
right; when GDB evaluates expressions NaN == NaN should be zero.
Therefore I think we shouldn't use this function when establishing
when a variable has been changed.  Does it make sense to simply do a
bit-for-bit comparison in that case?

Mark

Re: Bug in valarith.c:value_equal()?

[Andrew Cagney]

Mark Kettenis wrote:

Your right, varobj should use bitwise comparison.  Suggest a function 
value_content_equal.

Andrew

> I've found the cause of the testsuite problems I reported yesterday.
> The additional testsuite failures are intermittent.  If you look
> careful at the gdb.mi/mi-var-cmd.exp test you'll see that the test is
> checking whether some uninitialized local variables have been changed.
> The testsuite failures indicate that sometimes, the floating-point
> variables change unexpectedly.  Some further investigation showed that
> these unexpected changes happened when the (unitialized) variables
> were NaNs.  All of a sudden things make sense.  The variables don't
> really change.  GDB tries to determine whether a variable changes by
> comparing its current value to a previous value.  This is done by
> calling valarith.c:value_equal().  For floating-point variables, this
> function does the following check:
> 
>     return value_as_double (arg1) == value_as_double (arg2);
> 
> Now in C this will return 0, if ARG1 and ARG2 are NaN, even if they
> are bit for bit equal.
> 
> Actually I think the implementation of valarithm.c:value_equal() is
> right; when GDB evaluates expressions NaN == NaN should be zero.
> Therefore I think we shouldn't use this function when establishing
> when a variable has been changed.  Does it make sense to simply do a
> bit-for-bit comparison in that case?
> 
> Mark
> 
> 
> 

Re: Bug in valarith.c:value_equal()?

[M.M. Kettenis]

Andrew Cagney <cagney@gnu.org> wrote:

> Mark Kettenis wrote:
>
> Your right, varobj should use bitwise comparison.  Suggest a function
> value_content_equal.

Consider it done.  The varobj stuff does a lot of dancing to catch errors.  I'll see if I can simplify that a bit, so it might take a few days.

Mark