Assignment of -0 in arithmetic expansion doesn't remove negative sign

[laurenthuberdeau]

Assignment of -0 in arithmetic expansion assigns -0 to variables on ARM. ksh on x86 and other shells remove the negative sign, which makes me think it may be due to some undefined behavior like in #770.

This short example shows the bug.

> ksh
$ : $(( a = 0 ))
$ : $(( b = -0 ))
$ : $(( c = +0 ))
$ echo $a # print 0
0
$ echo $b # print -0 instead of 0
-0
$ echo $c # print 0 as expected
0

[ko1nksm]

Here is another example on Linux on x86_64. I don't think this includes undefined behavior.

$ ksh -c 'v=0; (( v = -1 * v )); echo $v'
-0

This issue may have been introduced between ksh93-2007-06-28 and ksh93-2007-11-05.

root@6eb222d646da:/# ksh93-2007-06-28 -c 'v=0; (( v = -1 * v )); echo $v'
0

root@6eb222d646da:/# ksh93-2007-11-05 -c 'v=0; (( v = -1 * v )); echo $v'
-0

[McDutchie]

My stash of historic versions (compiled on Debian arm64) says it was introduced between 2007-09-26 and 2007-10-31:

$ ksh_2007-09-26 -c 'v=0; (( v = -1 * v )); echo $v'
0
$ ksh_2007-10-31 -c 'v=0; (( v = -1 * v )); echo $v'
-0

[McDutchie]

There was only one arithmetic-related change between 2007-09-26 and 2007-10-31, and it's this one:

    Version: 2007-10-31 ksh93s+
    
    Sourced using the following ksh93-integration diff:
    3721fc2fa6fe51d5060dae7054b5c2692c811da692925ce15942c4d083a5ba18  ksh93_20070922_to_20071031.diff.txt

diff --git a/src/cmd/ksh93/sh/arith.c b/src/cmd/ksh93/sh/arith.c
index cb661177..f2e3773f 100644
--- a/src/cmd/ksh93/sh/arith.c
+++ b/src/cmd/ksh93/sh/arith.c
@@ -343,6 +343,8 @@ Sfdouble_t sh_strnum(register const char *str, char** ptr, int mode)
 		if(!ptr && *last && mode>0)
 			errormsg(SH_DICT,ERROR_exit(1),e_lexbadchar,*last,str);
 	}
+	else if (!d && *str=='-')
+		d = -0.0;
 	if(ptr)
 		*ptr = last;
 	return(d);

[McDutchie]

Here's a patch to revert that change on ksh 93u+m:

diff --git a/src/cmd/ksh93/sh/arith.c b/src/cmd/ksh93/sh/arith.c
index b53b0f36c..e32b5f08e 100644
--- a/src/cmd/ksh93/sh/arith.c
+++ b/src/cmd/ksh93/sh/arith.c
@@ -594,8 +594,6 @@ Sfdouble_t sh_strnum(const char *str, char** ptr, int mode)
 				}
 			}
 		}
-		else if(!d && *str=='-')
-			d = -0.0;
 	}
 	if(ptr)
 		*ptr = last;

Removing those lines fixes this issue (as expected), but introduces two regression test failures:

	arith.sh[504]: FAIL: %g of x $x $((x)) for x=-0 should all be -0
	arith.sh[505]: FAIL: %g of y $y $((y)) for y=-0.0 should all be -0

Those tests are (see lines 499, 504, 505):

ksh/src/cmd/ksh93/tests/arith.sh

Lines 499 to 505 in ff73450

	float x=-0 y=-0.0
	r=-0
	[[ $((-0)) == 0 ]] || err_exit '$((-0)) should be 0'
	[[ $(( -1*0)) == 0 ]] || err_exit '$(( -1*0)) should be 0'
	[[ $(( -1.0*0)) == -0 ]] || err_exit '$(( -1.0*0)) should be -0'
	[[ $(printf "%g %g %g\n" x $x $((x)) ) == '-0 -0 -0' ]] || err_exit '%g of x $x $((x)) for x=-0 should all be -0'
	[[ $(printf "%g %g %g\n" y $x $((y)) ) == '-0 -0 -0' ]] || err_exit '%g of y $y $((y)) for y=-0.0 should all be -0'

[McDutchie]

I'm now leaning towards thinking that this is not a bug.

Zero is signed in floating point arithmetic; 0.0 and -0.0 are different. On ksh, floating point arithmetic is (currently) the default.

If you declare the variable as integer, the problem goes away:

$ typeset -i v; ((v = -0)); echo $v
0

Perhaps if a solution to #771 is ever figured out, this may change.