https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79720
Bug ID: 79720
Summary: floating point result depends on optimization level
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: tkoenig at gcc dot gnu.org
Target Milestone: ---
The following test case
$ cat recip.f90
module try
implicit none
contains
complex function r1(a)
complex, value :: a
r1 = 1./a;
end function r1
complex function r2(a)
complex, value :: a
real :: a_real, a_imag;
real :: v_6, v_7, v_9, v_10, v_11, v_12, v_13, v_14, v_15, v_16, v_17, &
& v_18, v_19, v_20, v_21, v_22, v_23, v_24, v_27, v_28, &
& v_29, v_30, v_31, v_32
logical :: v_8
a_real = real(a)
a_imag = aimag(a)
v_6 = abs(a_real);
v_7 = abs(a_imag)
v_8 = v_6 < v_7;
if (.not.v_8) then
v_9 = a_real / a_imag;
v_10 = a_real * v_9;
v_11 = a_imag + v_10;
v_12 = v_9 + 0.0;
v_13 = v_9 * 0.0;
v_14 = v_13 - 1.0e+0;
v_15 = v_12 / v_11;
v_16 = v_14 / v_11;
v_31 = v_15;
v_32 = v_16;
v_27 = v_31;
v_28 = v_32;
else
v_17 = a_imag / a_real;
v_18 = a_imag * v_17;
v_19 = a_real + v_18;
v_20 = v_17 * 0.0;
v_21 = v_20 + 1.0e+0;
v_22 = 0.0 - v_17;
v_23 = v_21 / v_19;
v_24 = v_22 / v_19;
v_29 = v_23;
v_30 = v_24;
v_27 = v_29;
v_28 = v_30;
end if
r2 = cmplx(v_27, v_28);
end function r2
end module try
program main
use try
implicit none
complex :: c, a1, a2
c = (1.2e10, -3.2)
a1 = r1(c)
a2 = r2(c)
print *,a1, a2
print *,a1-a2
end program main
gives different results without optimization and with -O on x86_64
with a reasonably current trunk:
ig25@linux-d6cw:~/Krempel/Complex> gfortran recip.f90 && ./a.out
(8.333333329E-11,2.222222170E-20) (8.333333329E-11,2.222222331E-20)
(0.00000000,-1.615587134E-27)
ig25@linux-d6cw:~/Krempel/Complex> gfortran -O recip.f90 && ./a.out
(8.333333329E-11,2.222222331E-20) (8.333333329E-11,2.222222331E-20)
(0.00000000,0.00000000)
It does not appear to be processor-specific, POWER7 exhibits the same
behavior:
[tkoenig@gcc1-power7 ~]$ gfortran recip.f90 && ./a.out
( 8.33333333E-11, 2.22222217E-20) ( 8.33333333E-11, 2.22222233E-20)
( 0.00000000 , -1.61558713E-27)
[tkoenig@gcc1-power7 ~]$ gfortran -O recip.f90 && ./a.out
( 8.33333333E-11, 2.22222233E-20) ( 8.33333333E-11, 2.22222233E-20)
( 0.00000000 , 0.00000000 )
In the absence of -funsafe-math-optimizations, this is not supposed
to happen.
I would assume that this is not a dup of PR 72824; at least
-fno-tree-loop-distribute-patterns has no effect.
