draine@astrovax.UUCP (Bruce Draine) (02/25/84)
Subject: Bug in 4.2BSD f77 complex arithmetic
Description:
Incorrect evaluation of complex arithmetic expressions
involving library functions CCOS and CSIN. For example,
the fortran expression
A=CCOS(Y)/CSIN(Y)
with A and Y having been previously "typed" by a
COMPLEX A,Y
statement, is not being evaluated correctly for arbitrary
values of the complex variable Y.
It is not known to what extent this problem may also be
endemic to other constructions [e.g., A=CSIN(Y)/CCOS(Y) ].
Repeat-By:
Compiling and running the following simple program:
C PROGRAM TO TEST APPARENT BUG IN F77 COMPILER
C B.T.DRAINE,PRINCETON UNIV. OBS., 84/2/19
COMPLEX Y,A,B,C
Y=(4.601765,-3.839218)
WRITE(6,9900)Y
9900 FORMAT(' Y=',1P2E14.7)
C INDIRECT EVALUATION OF COTANGENT(Y)
A=CCOS(Y)
B=CSIN(Y)
C=A/B
WRITE(6,9969)A,B
9969 FORMAT(' A=CCOS(Y)=',1P2E14.7,' B=CSIN(Y)=',2E14.7)
WRITE(6,9970)C
9970 FORMAT(' C=A/B=',1P2E14.7)
C DIRECT EVALUATION
C=CCOS(Y)/CSIN(Y)
WRITE(6,9971)C
9971 FORMAT(' C=CCOS(Y)/CSIN(Y)=',1P2E14.7)
STOP
END
The following output is generated by this program running under
4.2 BSD on our Vax-11/750 with FPA:
***** testf77 compiled with no options 84/2/19 *****
output follows:
Y= 4.6017652e+00-3.8392179e+00
A=CCOS(Y)=-2.5673470e+00-2.3091778e+01 B=CSIN(Y)=-2.3113155e+01 2.5649724e+00
C=A/B= 2.0288443e-04 9.9909759e-01
C=CCOS(Y)/CSIN(Y)= 1.0000000e+00-1.4885653e-09
***** testf77 compiled with -C option 84/2/19 *****
output follows:
Y= 4.6017652e+00-3.8392179e+00
A=CCOS(Y)=-2.5673470e+00-2.3091778e+01 B=CSIN(Y)=-2.3113155e+01 2.5649724e+00
C=A/B= 2.0288443e-04 9.9909759e-01
C=CCOS(Y)/CSIN(Y)= 1.0000000e+00-1.4885653e-09
Fix:
Unknown to me. Has been reported to ucbvax\!bsd-bugs.ray@utcsrgv.UUCP (Raymond Allen) (02/27/84)
This is not a bug, FORTRAN protocol insists that the the functions as well as the variables have to be typed, so you need another declaration: COMPLEX CSIN,CCOS added to your program (otherwise FORTRAN assumes that CSIN and CCOS return REAL data type. You might also use the -u option to f77. With this option, any variable or function appearing in the program that is not defined in a type definition statement causes a fatal error. (Note: I did not actually run the corrected program but we use lotsa complex fortran here and it seems to work all right.) from the FORTRAN infested keyboard of: Ray Allen utcsrgv!ray (416) 978-5036
pam@lanl-a.UUCP (08/02/84)
Can anyone provide a fix to this bug in f77 on 4.2BSD? Please reply
directly to me, and I'll summarize to net.lang.f77. The bug does not
occur if b=cmplx(abs(real(a)),abs(imag(a))) is replaced with
b=cmplx(real(a),abs(imag(a))).
complex a, b
a = (1.0, 2.0)
b = cmplx (abs(real(a)), abs(imag(a)))
print *, a, b
stop
end
.data 0
.align 2
L15:
.long 0x4080
.long 0x4100
.align 2
L22:
.long 0x0
.align 2
L23:
.long 6
.align 2
L24:
.long 1
.align 2
L27:
.byte 00,00
.set LWM1,0x800
.data 2
.data 1
.data 0
.globl _MAIN_
.globl _r_imag
.globl _s_wsle
.globl _do_lio
.globl _e_wsle
.globl _s_stop
.align 2
.stabs "complex.f",0x64,0,0,0
.text
.globl _MAIN_
_MAIN_:
.word LWM1
subl2 $LF1,sp
jmp L12
L13:
movf L15,v.2-v.1(r11)
movf L15+4,v.2-v.1+4(r11)
movf v.2-v.1(r11),-4(fp)
cvtfd -4(fp),-24(fp)
cmpf L22,-4(fp)
jgtr L9999
movf -4(fp),r0
jbr L9998
L9999:
mnegf -4(fp),r0
L9998:
cvtfd r0,-40(fp)
cvtdf -40(fp),-32(fp)
clrf -28(fp)
movf -32(fp),v.3-v.1(r11)
movf -28(fp),v.3-v.1+4(r11)
pushl $v.4
calls $1,_s_wsle
pushl $8
addl3 $v.2-v.1,r11,r0
pushl r0
pushl $L24
pushl $L23
calls $4,_do_lio
pushl $8
addl3 $v.3-v.1,r11,r0
pushl r0
pushl $L24
pushl $L23
calls $4,_do_lio
calls $0,_e_wsle
pushl $0
pushl $L27
calls $2,_s_stop
L14:
ret
.align 1
L12:
movl $v.1,r11
jmp L13
.set LF1,40
.text
.data 2
.align 2
v.4:
.long 0
.long 6
.space 4
.long 0
.align 2
v.2:
.space 8
.align 2
v.3:
.space 8
.set v.1,v.3
output from executable:
( 1.00000, 2.00000) ( 1.00000, 0.)
The above is incorrect. The correct output should be:
( 1.00000, 2.00000) ( 1.00000, 2.00000)