I would do several things in this situation:
1. Download the Intel (AMD?) manuals.
2. Read the parts that deal with x87 exceptions.
3. Re-compile the code with no optimization, obtaining the assembly code
(-S option), and see if the problem still exists.
4. Repeat steps 2 and 3, adding in more optimization, as needed.
I note the following from "Intel(R) 64 and IA-32 Architectures Software
Developer's Manual":
8.5.5 Numeric Underflow Exception (#U)
The x87 FPU detects a floating-point numeric underflow condition
whenever the
rounded result of an arithmetic instruction is tiny; that is, less than
the smallest
possible normalized, finite value that will fit into the floating-point
format of the
destination operand. (See Section 4.9.1.5, “Numeric Underflow Exception
(#U),” for
additional information about the numeric underflow exception.)
--Bob
On 8/9/2011 9:01 AM, Olivier Maury wrote:
Ok there is one main difference when the value to be put on the stack
is NAN : it is not put on the stack !
The fldl instruction just do nothing (expected ?) and that's why
instead of having two values on the FP stack I have only one value !
Even if the intel documentation says nothing about that configuration
I guess that's why the fcmove raise a FPE...
What do you think ?
On 08/09/11 16:44, Ian Lance Taylor wrote:
Olivier Maury<Olivier_maury@xxxxxxxxxx> writes:
I was wondering, is the following assembly code correct:
c94: e8 fc ff ff ff call c95
<load_param_def_given+0x1d1>
c99: 84 c0 test %al,%al
c9b: dd 45 d0 fldl -0x30(%ebp)
c9e: dd 45 d8 fldl -0x28(%ebp)
ca1: da c9 fcmove %st(1),%st
ca3: dd d9 fstp %st(1)
ca5: dd 5d d0 fstpl -0x30(%ebp)
ca8: 89 f8 mov %edi,%eax
This code is some optimized assembly code extracted from the object
file using objdump. My concern here is that I have a test instruction
without the corresponding jump !
The test instruction is setting the flags for the fcmove instruction.
That piece of assembly comes from a code that look like:
if (my_function(param1,&out_param))
value = out_param;
with :
out_param a double value that is not assigned with a default value
and
char my_function(long param1, double *out)
{
char ret = 0;
... do some stuff ...
if (some_property)
{
... do some stuff ...
*out = a_value_computed;
ret = 1;
}
....
return ret;
}
And taking that optimized code in a debugger it crash with a FPE from
time to time because even if the my_function returns 0 it seems to do
the assignement (value = out_param) with an unitialized out_param
value !
What do you think ? Am I doing something wrong or is it a gcc bug ?
It's pretty hard to say anything about an incomplete example. However,
fcmove is a conditional move, so the code will only do the move if the
value returned is non-zero. I don't see anything wrong in what you have
shown us.
Ian
