Both arm and arm64 have rules like:

(SUBshiftLL x (SLLconst x [c]) [d]) && c==d => (MOVDconst [0])

This actually encodes the identity x - x<<c<<c == 0, which is incorrect. It should be

(SUBshiftLL (SLLconst x [c]) x [d]) && c==d => (MOVDconst [0])

Which encodes the correct identity x<<c - x<<c == 0.

It turns out to be quite hard to trigger this bad rule, as x<<c<<c is almost always reduced by the generic rewrite pass. But here's a case which gets through:

package main

func f(x uint64) uint64 {
	s := "\x05"
	c := s[0]
	return x - x<<c<<5
}

func main() {
	println(f(5))
}

This should print some giant number (18446744073709546501), but on arm64 it prints 0.

@cherryyz