A slight modification of El Gamal signature scheme would be to compute:
$r = g^k$ for some random k (as usual) but $s$ as:
$s = (hash(m) - x)k^{-1} \mod (p-1) $ (instead of $(hash(m)-xr)k^{-1} \mod p - 1$).
The verification procedure would be to check:
$g^{hash(m)} = g^x r^s \mod p$
I don't see what we are losing by doing so.
Forging an arbitrary ElGamal signature means finding $(r,s)$ such that $g^h = y^r r^s \mod p$ for given values of $p$, $g$, $h$ and $y$. We don't know how to do that other than selecting $r$ and then having to solve a discrete logarithm for $s$.
In your proposed variant, forging a signature means finding $(r,s)$ such that $g^h = y r^s$. This is just a matter of taking $s = 1$ and $r = g^h / y$. This can't be fixed by adding another constraint on $s$ since the problem would then be to find an $n$th root, which is solvable.
The gist of ElGamal variants is to involve a linear combination of $r$ and $g^r$ during verification. This is what makes a forgery likely to be as hard as solving a discrete logarithm.
