Re: subscript expanded twice in a[subcript]++ (Was: problem with RANDOM and arrays)

[Bart Schaefer <schaefer@xxxxxxxxxxxxxxxx>]

On Jan 20,  1:32pm, Stephane Chazelas wrote:
}
} > My understanding is that
} > 
} > x[RANDOM%2+1]++
} > 
} > is handled as
} > 
} > x[RANDOM%2+1] = x[RANDOM%2] + 1

It's not quite that simple; if it were, this would also fail:

schaefer<519> for i ({1..1000}) ((x[$((RANDOM%2+1))]++)); echo $x
506 494

} > Similar problem:
} > 
} > $ x=0; ((a[++x]++)); echo $x
} > 2

I find a slightly different problem:

schaefer<520> unset x a; repeat 2 { x=0; ((a[++x]++)); print X $x A $a }
X 1 A 1
X 2 A 1 2

Note that the first time around, the op works as expected.  This seems
to have to do with whether $a is unset:

schaefer<530> unset x a; declare -a a
schaefer<531> repeat 2 { x=0; ((a[++x]++)); print X $x A $a }
X 2 A 1
X 2 A 1

However, the error *does* seem to be caused by both getnparam() and
setnparam() calling getvalue(), which eventually calls getindex() where
the subscript expression is parsed and evaluated by matheval().

The calls are at math.c:1113 and math.c:1137 in the specific case of
the "++" operator, but of course every op has its own setvar() call.

Interestingly the problem for zsh seems to be that array elements are
always scalar, never integer, so they must always be converted back
and forth to strings when arithmetic is performed on them.  (I can't
get to the ksh list archive to see if the problem there was similar.)

I don't see any simple fix for this.  We could parse a[++x] as if it
were a[$((++x))] before entering math context, but then ternaries
((x ? a[x++] : a[++x])) would incorrectly increment in both branches.
Anyway that's problematic as I don't believe the parser knows whether
$a references a scalar array or a hash.

Another option is a completely new set of APIs for fetching and
setting variable values, which would manipulate a handle to the internal
storage of a[N] rather than returning the value stored there so that
we could both read and update through the same handle.  This almost
exists e.g. getvalue()/getnumvalue() and setnumvalue(), but there are
potential recursion issues, e.g., getnumvalue() also may make a call
to matheval() because of this:

schaefer<549> a=('y+2'); y=5; print $((++a[1]))
8

I'd like to know if ksh supports that and if so how it avoids side-
effects on evaluating the string value of a[1] as a math expression.

schaefer<550> x=0; a=(x+=3 x+=2 x+=1 x=0); ((++a[++x])); print $a
x+=3 x+=2 x+=1 x=0 5