I am aware of the 80bit float limitation of valgrind, but I am confused by the 
inconsistent behaviour shown by the test case following. As I understand the 
documentation, 80 bit floats are treated as if they were 64 bits, so generally 
the results should be the same provided that the numbers do not require high 
precision.

The following testcase is Ada code, compiled with GNAT pro 5.05 on an uptodate 
debian unstable. The command to compile is "gnatmake main". The source file 
should be called main.adb. The output without valgrind is:

./main
 1.59990000000000000E+01
 1.60000000000000000E+01
 1.60000000000000000E+01

valgrind ./main
(cut)
 1.59990000000000000E+01
 1.60000000000000000E+01
 1.00000000000000000E+00

It seems that the conversion to 64 bit works correctly with a constant but not 
with a local variable.

File: main.adb


with text_io; use text_io;
procedure main is 
     type simple_float is new long_long_float;

     g : simple_float;
     f : simple_float;
     h : simple_float;
   begin
     g := 15.999;
     f := simple_float'ceiling(simple_float(15.999));
     h := simple_float'ceiling(g);

     put_line( simple_float'image(g));
     put_line( simple_float'image(f));
     put_line( simple_float'image(h));

end main;