GCC from SuSE8.2 has bug in handling of floating point constants

[Joerg Wleklik <joerg at wleklik dot de>]

Hi,

I detected an bug in the handling of floating point constants in the GCC 3.3 
version, delivered with SuSE Linux 8.2.

If there is not explicit an other definition an floating point constant has to 
be handled as a double value. The GCC version from the SuSE 8.2 does not so.
If the value of the constant is in the range of float the compiler handles the 
value as float value. That seems to be OK but if you compare a double 
variable against the constant, the double variable will be converted into 
float and that may fail.

Example:

A simple if-statement in 'C'

	if(Block > 999999.)

will be compiled to:

      fldl     0xfffffff8(%ebp)     <--     Load the value of Block
      fstps  0xffffffec(%ebp)    <--     Store the value and convert
                                                    convert it float
      flds    0xffffffec(%ebp)    <--     Load value of Block
      flds    0x815a51c           <--     Load the constant
      fxch   %st(1)                  <--     Load value of st(1) to st(0)
      fucompp                        <--     Compare the values
        :               :
        :               :

If the value of Block is bigger as the range of 'float' this code results in 
an FPU-OVERFLOW exception.

The official GCC 3.3 version from gnu.org compiles the if-statement to:

      fldl     0xfffffff8(%ebp)     <--     Load the value of Block
      fldl     0x815a51c           <--     Load the constant
      fxch   %st(1)                  <--     Load value of st(1) to st(0)
      fucompp                        <--     Compare the values
       :               :
       :               :
This works as expected without an exception.


There is a short example attached.

Regards,
Joerg Wleklik

Attachment: main.cpp.tgz
Description: application/tgz