[Bug c++/13719] New: 0.2*5.0>1.0 even with -O0, missing fldl probably

[guilhem at mysql dot com]

Hi!
I have read the "most reported non-bug" about floating-point precision, but I
think it's a real bug here; please take time to read.
I personally use 3.2.2 but colleague with 3.3.2 has the same bug too (with
same testcase and same compiler options). We use Linux (Mandrake & SuSE).

Here is test.cc:
#include <stdlib.h>

int main()
{
  return ((int)((atof("0.2")*atof("5.0"))>1.0));
}

Compile this with:
g++ -O0    -o test_bin test.cc
And in a shell execute:
./test_bin ; echo $?
I get 1, so comparison is done wrong.
Adding -ffloat-store does not remove the error.

Here is the problem in my humble opinion (using the output of g++ -S):
g++ -O0 -S test.cc:

        .file   "test.cc"
        .section        .rodata
.LC0:
        .string "0.2"
.LC1:
        .string "5.0"
        .text
        .align 2
.globl main
        .type   main,@function
main:
.LFB1:
        pushl   %ebp
.LCFI0:
        movl    %esp, %ebp
.LCFI1:
        subl    $8, %esp
.LCFI2:
        andl    $-16, %esp
        movl    $0, %eax
        subl    %eax, %esp
        subl    $12, %esp
        pushl   $.LC0
.LCFI3:
        call    atof
        addl    $16, %esp
        fstpl   -8(%ebp)
        subl    $12, %esp
        pushl   $.LC1
        call    atof
        addl    $16, %esp
        fldl    -8(%ebp)
**multiplication of the 2 numbers
        fmulp   %st, %st(1)
**result of multiplication (1.000000000000000056)
**is left in FPU
**and directly used for comparison, hence the problem
        fld1
        fxch    %st(1)
        fucompp
        fnstsw  %ax
        testb   $69, %ah
        sete    %al
        andl    $255, %eax
        leave
        ret
.LFE1:
.Lfe1:
        .size   main,.Lfe1-main
        .ident  "GCC: (GNU) 3.2.2 (Mandrake Linux 9.1 3.2.2-3mdk)"

I guess there should be some rounding (using fldl) between multiplication and
comparison; it does not make sense to do the comparison with the extra precision
of the FPU. In the next example we see it better.

If I add a few printf(), the error appears only with -O1 (or more): here is
test2.cc:
#include <stdlib.h>
#include <stdio.h>

main()
{
  double e = atof("0.2")*atof("5.0");
  if (e>1.0)
    printf("0.2 * 5.0 > 1.0 (ERROR)\n");
  else
    printf("0.2 * 5.0 <= 1.0 (OK)\n");
}

When built with g++ -O0 test2.cc it prints OK, with -01 it prints ERROR.
Here are snippets of the assembler code (the multiplication and comparison section):
with -O0:
<cut>
        fmull   -16(%ebp)
        fstpl   -8(%ebp)
        fldl    -8(%ebp)
        fld1
        fxch    %st(1)
        fucompp
<cut>
(see fldl which does the rounding of the multiplication result from FPU
precision to double precision).
With -O1:
<cut>
        fmulp   %st, %st(1)
        fld1
        fxch    %st(1)
        fucompp
<cut>
Again, the result of the multiplication (with precision of the FPU) is used
directly for comparison, without any rounding.

In the MySQL code which is were the problem originally arose (we had
SELECT ACOS(0.2*5.0) returning NULL), the problem appeared only with -O2 (or more).

I hope these testcases can be of use.
Guilhem Bichot (MySQL A.B.)

-- 
           Summary: 0.2*5.0>1.0 even with -O0, missing fldl probably
           Product: gcc
           Version: 3.3.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: c++
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: guilhem at mysql dot com
                CC: gcc-bugs at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13719